From Question-First to Analyst-First: Domain-Expert Skills and Verified Knowledge Compilation for Proactive Enterprise Analytics

# Domain-Expert Skills and Verified Knowledge Compilation for Proactive Enterprise Analytics
Source: [https://arxiv.org/html/2608.28594](https://arxiv.org/html/2608.28594)
## From Question\-First to Analyst\-First: Domain\-Expert Skills and Verified Knowledge Compilation for Proactive Enterprise Analytics

\(June 15, 2026\)

###### Abstract

Conversational analytics systems overwhelmingly assume the user already has a well\-formed question, leaving a non\-expert facing a blank query box on an unfamiliar enterprise schema\. Commercial “proactive” tools narrow this gap only by detecting statistical anomalies over analyst\-curated metric layers, and academic next\-question recommenders depend on query logs that a fresh dataset lacks\. We describe a production enterprise analytics system that inverts the interaction model from*question\-first*to*analyst\-first*through two coupled architectural ideas\. First, a pluggable domain\-expert*skill*abstraction: a folder\-based, database\-free subject\-matter pack \(a manifest, per\-stage prompt facets, keyword\-routed reference libraries, standing\-report templates, and optional deterministic compute\) that is auto\-selected per \(client, dataset\) by a deterministic schema\-matching function and spliced as a*cross\-cutting concern*into every stage of a streaming agentic pipeline, the schema explorer, and the report engines, degrading to a strict no\-op when absent\. Because a skill is a self\-contained folder resolved deterministically and added with no per\-client engineering, the catalogue is open\-ended: an extensible*marketplace*of domain experts rather than a fixed feature set\. Second, an offline*knowledge\-compilation*loop: a Claude\-Code\-style agent probes the dataset’s parquet via DuckDB \(zero load on production\), runs critic\-gated per\-table convergence with self\-healing retries, and data\-validates joins by value overlap, producing durable schema knowledge that, together with the resolved skill, drives standing expert reports whose every published metric is re\-verified by independently re\-executing its evidence SQL, and suggested questions that mirror the report agenda\. These close a*proactive loop*: reports surface numbers, the numbers seed questions, and a click launches a verified deep dive, all before the query box is ever used\. We give a formal model of the injection operator and the compilation function, describe the fully\-automatic onboarding chain, and report illustrative evidence from a single live tenant: deep exploration over10/1010/10tables with2222data\-validated joins, seven chartered\-accountant standing reports, and roughly thirty domain questions generated with no prior user query\. We make no user\-study or benchmark claims; the contribution is the architecture and its defensibility\.

## 1Introduction

The dominant interface to enterprise data is now a text box\. A decade of work on natural\-language interfaces to databases and, more recently, large\-language\-model \(LLM\) text\-to\-SQL has driven translation accuracy upward on academic benchmarks\[[60](https://arxiv.org/html/2608.28594#bib.bib1),[30](https://arxiv.org/html/2608.28594#bib.bib2),[22](https://arxiv.org/html/2608.28594#bib.bib3)\]and into commercial products\[[43](https://arxiv.org/html/2608.28594#bib.bib61),[9](https://arxiv.org/html/2608.28594#bib.bib63)\]\. Yet every one of these systems shares a hidden premise:*the user already knows what to ask*\. A perfect translator from English to SQL does nothing for an analyst who is staring at a freshly connected schema of eighty opaque tables and does not know which question is worth a query\. This is the*cold\-start*or*blank\-query\-box*problem, and it is a documented barrier to self\-service business\-intelligence adoption\[[27](https://arxiv.org/html/2608.28594#bib.bib10)\]\. The hard part of analysis is not phrasing a question; it is knowing which question a domain expert would ask of*this*data\.

Two families of systems try to supply the question rather than merely answer it, and each cold\-starts in its own way\.*Log\-based next\-question recommenders*mine the queries of prior users to suggest the next step\[[6](https://arxiv.org/html/2608.28594#bib.bib7),[54](https://arxiv.org/html/2608.28594#bib.bib8)\], but a new tenant has no logs, so the recommender itself has nothing to recommend from on day one\.*Automated insight discovery*sweeps a dataset for statistically salient patterns\[[15](https://arxiv.org/html/2608.28594#bib.bib9),[13](https://arxiv.org/html/2608.28594#bib.bib27),[52](https://arxiv.org/html/2608.28594#bib.bib23)\], but “statistically interesting” is not the same as “what a chief financial officer or a chartered accountant would act on,” and these tools are domain\-agnostic by construction\. The most proactive commercial offerings divide into two camps that share the same blind spot\. Tableau Pulse is bound to an*analyst\-curated metric layer*: a human first defines the metrics and the system narrates their movements\[[1](https://arxiv.org/html/2608.28594#bib.bib58),[2](https://arxiv.org/html/2608.28594#bib.bib57)\]\. ThoughtSpot SpotIQ and Power BI Quick Insights are domain\-agnostic statistical insight miners: anomaly, correlation, and trend detection over a dataset\[[49](https://arxiv.org/html/2608.28594#bib.bib59),[37](https://arxiv.org/html/2608.28594#bib.bib60)\]\. Neither camp carries a*portable domain expert*that, given a matching skill for the domain, can approach an un\-curated schema and author the agenda itself\.

#### Thesis\.

This paper describes a system that flips the interaction model from question\-first to analyst\-first by treating domain expertise as a*first\-class, pluggable, auto\-selected software artifact*and by*compiling*durable, verified knowledge about a dataset*before*the user asks anything\. The two ideas are coupled\. A domain\-expert*skill*is a self\-contained folder \(prompt fragments, reference libraries, standing\-report templates, deterministic compute\) that the system splices into every node of its agentic pipeline and offline engines\. An offline*knowledge\-compilation*loop, modeled on the way a coding agent explores a codebase, probes the data until an independent critic is satisfied, validates joins against the data, and persists a reusable knowledge base\. The skill and the knowledge together drive \(i\) standing expert reports whose every published number is re\-verified by re\-executing its evidence SQL, and \(ii\) suggested questions drawn from the skill’s review\-cadence playbook and grounded in the validated knowledge\. The result is a closed*proactive loop*\([Fig\.˜4](https://arxiv.org/html/2608.28594#S7.F4)\): the system surfaces*numbers*the user did not ask for, the numbers*seed questions*, and a click launches a verified*deep dive*, a sequence that runs to completion before the blank box is ever touched\.

#### Why the fusion, not the parts, is the contribution\.

Each ingredient in isolation is incremental\. Skill packaging echoes the modular “agent skills” and subagent patterns now emerging as general developer tooling\[[4](https://arxiv.org/html/2608.28594#bib.bib41),[5](https://arxiv.org/html/2608.28594#bib.bib43)\]; agentic exploration and critic\-gated refinement are established\[[59](https://arxiv.org/html/2608.28594#bib.bib11),[47](https://arxiv.org/html/2608.28594#bib.bib12),[18](https://arxiv.org/html/2608.28594#bib.bib50)\]; execution\-grounded verification has precedent\[[18](https://arxiv.org/html/2608.28594#bib.bib50),[8](https://arxiv.org/html/2608.28594#bib.bib51)\]\. The contribution is the*fusion*, deployed: a per\-tenant, schema\-auto\-selected domain\-expert layer threaded as a cross\-cutting concern through an entire analytics stack, fused with a data\-validated knowledge base and a report engine in which a number that fails to reproduce against its own evidence SQL is suppressed rather than published, instantiated as a closed proactive loop with a fully\-automatic onboarding chain\. We return to this argument in[Section˜10](https://arxiv.org/html/2608.28594#S10)\.

#### Scope and honesty\.

This is a single\-deployment*system*paper\. We report no controlled user study, no A/B test, and no public\-benchmark leaderboard; our deployment figures \([Section˜9](https://arxiv.org/html/2608.28594#S9)\) are illustrative observations from one live tenant, reported for transparency rather than as measured wins\. Where the architecture makes a guarantee, we state its boundary precisely: the verification layer certifies that a published number*matches its declared evidence SQL*within tolerance, not that the metric’s business definition is semantically correct\.

#### Contributions\.

1. 1\.A pluggable, database\-freedomain\-expert skill abstraction\([Section˜4](https://arxiv.org/html/2608.28594#S4)\) that injects opinionated expertise as a cross\-cutting concern at every node of an agentic analytics pipeline, the schema explorer, and both report engines, while degrading to an exact no\-op when no skill is active\.
2. 2\.A deterministic, source\-agnosticskill\-selection algorithmthat resolves one active skill per \(client, dataset\) from schema structure, including distinctive*cell values*that let expert framing fire on opaque schemas \(SAP T\-tables, Tally exports\) where names alone fail, with no query logs and no per\-client engineering\.
3. 3\.An offlineknowledge\-compilation loop\([Section˜5](https://arxiv.org/html/2608.28594#S5)\): a JSON\-action probe agent over read\-only DuckDB\-on\-parquet, gated by a fast deterministic coverage critic, wrapped in a self\-healing convergence loop whose work list provably shrinks, that data\-validates joins by value overlap and persists durable, resumable schema knowledge\.
4. 4\.Anexecution\-grounded report\-synthesis engine\([Section˜6](https://arxiv.org/html/2608.28594#S6)\) in which every published metric carries an evidence query that an independent, non\-LLM critic re\-executes and tolerance\-matches, so a number that does not reproduce against its evidence query is suppressed rather than shown, and a durable recipe drives zero\-LLM, zero\-kernel refresh and drill\-down\.
5. 5\.Theproactive\-analyst loop and its fully\-automatic onboarding chain\([Section˜7](https://arxiv.org/html/2608.28594#S7)\), in which the*same*compiled knowledge and skill power reports, questions, and live chat\.
6. 6\.Aformal model\([Section˜3](https://arxiv.org/html/2608.28594#S3)\) of the injection operator \(with a no\-op/backward\-compatibility property\) and the compilation function \(with a termination argument\), plus an honest single\-deployment account \([Sections˜9](https://arxiv.org/html/2608.28594#S9)and[11](https://arxiv.org/html/2608.28594#S11)\)\.

## 2Related Work

#### Text\-to\-SQL and the translation framing\.

Semantic parsing of natural language to SQL has matured from cross\-domain benchmarks\[[60](https://arxiv.org/html/2608.28594#bib.bib1)\]to knowledge\-grounded, value\-sensitive benchmarks\[[30](https://arxiv.org/html/2608.28594#bib.bib2)\]and broad LLM\-era surveys\[[22](https://arxiv.org/html/2608.28594#bib.bib3),[46](https://arxiv.org/html/2608.28594#bib.bib38)\]\. Crucially,*Spider 2\.0*reframes the task around*real enterprise*workflows \(dozens of tables, dialects, and external knowledge\) and shows that translation alone is far from sufficient\[[26](https://arxiv.org/html/2608.28594#bib.bib20)\], motivating both schema scouting and an external knowledge layer of the kind we compile\. All of this work, however, takes the question as given\.

#### Disambiguating a stated question vs\. the prior question of what to ask\.

A complementary line helps users refine a question they have already posed: NaLIR interactively resolves parse ambiguity\[[29](https://arxiv.org/html/2608.28594#bib.bib4)\], while DataTone and Eviza manage ambiguity in natural\-language visualization interfaces\[[16](https://arxiv.org/html/2608.28594#bib.bib5),[45](https://arxiv.org/html/2608.28594#bib.bib6)\]\. These address*how to phrase*; we address*what to ask in the first place*\.

#### Cold\-start and next\-question recommendation\.

QueRIE recommends queries from collaborative analysis of prior sessions\[[6](https://arxiv.org/html/2608.28594#bib.bib7)\], andWanget al\.\[[54](https://arxiv.org/html/2608.28594#bib.bib8)\]recommend the next natural\-language step in an interactive session\. Both are log\- or session\-dependent and therefore cold\-start on a fresh dataset, precisely the gap a portable domain\-expert prior closes\. Recent work generates questions directly from tabular data for conversational exploration\[[7](https://arxiv.org/html/2608.28594#bib.bib30)\], the closest prior art to our question generator\. It derives questions from table statistics and content; ours differ in two ways that the rest of this paper makes concrete\. First, the agenda is seeded by the resolved skill’s KPI catalogue and review\-cadence playbook \([Section˜7](https://arxiv.org/html/2608.28594#S7)\), so the questions reflect what a role would ask rather than what is statistically describable\. Second, they are generated against the compiled, data\-validated schema \([Section˜5](https://arxiv.org/html/2608.28594#S5)\), so a suggested question is answerable by the same joins the live pipeline will use, an answerability constraint statistics\-only generation does not impose\.

#### Automated insight discovery and mixed\-initiative analytics\.

Mixed\-initiative interaction is a long\-standing ideal\[[23](https://arxiv.org/html/2608.28594#bib.bib22)\]\. Visualization recommenders\[[52](https://arxiv.org/html/2608.28594#bib.bib23),[56](https://arxiv.org/html/2608.28594#bib.bib24),[57](https://arxiv.org/html/2608.28594#bib.bib25),[48](https://arxiv.org/html/2608.28594#bib.bib26)\]and insight miners\[[15](https://arxiv.org/html/2608.28594#bib.bib9),[13](https://arxiv.org/html/2608.28594#bib.bib27),[34](https://arxiv.org/html/2608.28594#bib.bib28),[11](https://arxiv.org/html/2608.28594#bib.bib29)\]proactively surface salient patterns, and recent LLM agents push toward proactive assistance\[[62](https://arxiv.org/html/2608.28594#bib.bib31),[33](https://arxiv.org/html/2608.28594#bib.bib32)\]amid a renewed survey interest in the automation–initiative balance\[[38](https://arxiv.org/html/2608.28594#bib.bib33)\]\. These systems optimize statistical interestingness; our agenda is set by a domain expert and is verified against the data\.

#### LLM data\-science and multi\-agent systems\.

Reasoning\-and\-acting and reflective agents\[[59](https://arxiv.org/html/2608.28594#bib.bib11),[47](https://arxiv.org/html/2608.28594#bib.bib12)\], multi\-agent frameworks\[[58](https://arxiv.org/html/2608.28594#bib.bib13),[20](https://arxiv.org/html/2608.28594#bib.bib14)\], and data\-science agents\[[19](https://arxiv.org/html/2608.28594#bib.bib15)\]together with their evaluation benchmarks\[[24](https://arxiv.org/html/2608.28594#bib.bib16)\]establish the agentic substrate we build on\. Visualization\- and dashboard\-generation agents\[[14](https://arxiv.org/html/2608.28594#bib.bib17),[35](https://arxiv.org/html/2608.28594#bib.bib18),[61](https://arxiv.org/html/2608.28594#bib.bib44)\]target the artifact; we target the*closed loop*that decides which artifacts to make and re\-verifies their numbers against evidence SQL\.

#### Domain\-knowledge injection\.

Retrieval\-augmented generation and tool use inject external knowledge into LLM inference\[[28](https://arxiv.org/html/2608.28594#bib.bib34),[17](https://arxiv.org/html/2608.28594#bib.bib35),[42](https://arxiv.org/html/2608.28594#bib.bib36)\]; in the SQL setting, Knowledge\-to\-SQL trains a single “data\-expert” model to supply knowledge*upstream*to improve SQL generation\[[21](https://arxiv.org/html/2608.28594#bib.bib37)\], in contrast to our*downstream*evidence\-SQL publish gate\. Semantic/metrics layers\[[12](https://arxiv.org/html/2608.28594#bib.bib39),[32](https://arxiv.org/html/2608.28594#bib.bib40)\]encode metric definitions but require human curation per organization\. Modular “agent skills” and subagents package capability and orchestration as general developer tooling\[[4](https://arxiv.org/html/2608.28594#bib.bib41),[5](https://arxiv.org/html/2608.28594#bib.bib43)\]\. We specialize skill packaging into a per\-tenant, schema\-auto\-selected*analytics*skill that threads the same expertise across an entire pipeline and both report engines, and we derive the agenda from raw schema rather than a pre\-curated metric layer\.

#### Verifier\-/critic\-gated loops and agent memory\.

Verifier ranking\[[10](https://arxiv.org/html/2608.28594#bib.bib45)\], process supervision\[[31](https://arxiv.org/html/2608.28594#bib.bib46)\], self\-consistency\[[55](https://arxiv.org/html/2608.28594#bib.bib47)\], and self\-refinement\[[36](https://arxiv.org/html/2608.28594#bib.bib48)\]improve reliability, whileHuanget al\.\[[25](https://arxiv.org/html/2608.28594#bib.bib49)\]caution that*intrinsic*self\-correction is unreliable, motivating*execution\-grounded*critics\[[18](https://arxiv.org/html/2608.28594#bib.bib50),[8](https://arxiv.org/html/2608.28594#bib.bib51)\]and LLM\-as\-judge evaluation\[[63](https://arxiv.org/html/2608.28594#bib.bib52)\]; seePanet al\.\[[40](https://arxiv.org/html/2608.28594#bib.bib55)\]for a survey of automated\-correction strategies\. Persistent agent memory and skill libraries\[[41](https://arxiv.org/html/2608.28594#bib.bib53),[39](https://arxiv.org/html/2608.28594#bib.bib54),[53](https://arxiv.org/html/2608.28594#bib.bib19)\]inform our durable knowledge base\. We also draw on the relational data\-profiling literature\[[3](https://arxiv.org/html/2608.28594#bib.bib56)\]for what an automated explorer should measure\. Our distinctive move is to fuse an*opinionated*expert \(known to risk domain\-correlated persona bias\[[51](https://arxiv.org/html/2608.28594#bib.bib21)\]\) with a*deterministic*execution check, so the opinion frames the analysis but cannot fabricate the number\.

#### Commercial proactive analytics\.

Question\-first natural\-language products\[[43](https://arxiv.org/html/2608.28594#bib.bib61),[44](https://arxiv.org/html/2608.28594#bib.bib62),[9](https://arxiv.org/html/2608.28594#bib.bib63)\], metric\-bound proactive products\[[1](https://arxiv.org/html/2608.28594#bib.bib58),[2](https://arxiv.org/html/2608.28594#bib.bib57),[50](https://arxiv.org/html/2608.28594#bib.bib64)\], and domain\-agnostic statistical insight miners\[[49](https://arxiv.org/html/2608.28594#bib.bib59),[37](https://arxiv.org/html/2608.28594#bib.bib60)\]define the landscape\. The white space they leave \(proactivity over an un\-curated schema, for a domain a skill covers, driven by a portable domain expert and backed by execution verification\) is the space this system occupies \([Table˜2](https://arxiv.org/html/2608.28594#S10.T2)\)\.

## 3Problem Framing and Formalization

We model an enterprise analytics platform that is*multi\-tenant*\(everything is scoped by a client identifier with no defaults\) and serves two regimes: a*reactive*live pipeline that answers a query, and a*proactive*offline phase that produces artifacts before any query\. The formalization makes precise \(a\) what a skill is and how it is injected, \(b\) what knowledge compilation computes and why it terminates, and \(c\) the admissibility contract that lets a report publish a number\.

### 3\.1Sets and objects

LetCCbe the set of clients andDDthe set of datasets\. For each pair\(c,d\)\(c,d\)the dataset has a schemaΣ​\(c,d\)=\(T,𝐶𝑜𝑙,𝑉𝑎𝑙\)\\Sigma\(c,d\)=\(T,\\mathit\{Col\},\\mathit\{Val\}\), whereTTis the bag of table names,𝐶𝑜𝑙\\mathit\{Col\}the bag of column names, and𝑉𝑎𝑙\\mathit\{Val\}a sampled bag of cell values from the top rows\. The live pipeline is an ordered tuple of stagesP=\(s1,…,s6\)=\(guard,resolve,scout,collate,execute,narrate\)P=\(s\_\{1\},\\dots,s\_\{6\}\)=\(\\textsf\{guard\},\\textsf\{resolve\},\\textsf\{scout\},\\textsf\{collate\},\\textsf\{execute\},\\textsf\{narrate\}\), each a function on a run contextxx, i\.e\.si:X→Xs\_\{i\}:X\\to X\.PPis the*maximal*stage sequence: the route resolved inresolve\(cache\_hit,simple, orcomplex\) may bypassscoutandcollate, but because injection is per\-stage the results below hold on whichever sub\-sequence executes\.

###### Definition 1\(Skill\)\.

A*skill*is a tuplek=\(𝑠𝑙𝑢𝑔,μ,F,R,κ,ρ,Y,L,𝑐𝑓𝑔\)k=\(\\mathit\{slug\},\\mu,F,R,\\kappa,\\rho,Y,L,\\mathit\{cfg\}\)whereμ\\muis the auto\-detect spec\(any\_table,any\_column,any\_value,keywords,min\_score\)\(\\textsf\{any\\\_table\},\\textsf\{any\\\_column\},\\textsf\{any\\\_value\},\\textsf\{keywords\},\\textsf\{min\\\_score\}\);F:N→Σ∗F:N\\to\\Sigma^\{\*\}maps a node name inN=\{router,scout,coder,chart,narrator,report\_planner,report\_analyst\}N=\\\{\\textsf\{router\},\\textsf\{scout\},\\textsf\{coder\},\\textsf\{chart\},\\textsf\{narrator\},\\textsf\{report\\\_planner\},\\textsf\{report\\\_analyst\}\\\}to its*facet*text \(ε\\varepsilonif absent\);R:stem→Σ∗R:\\textit\{stem\}\\to\\Sigma^\{\*\}are reference documents with a keyword mapκ:stem→2W\\kappa:\\textit\{stem\}\\to 2^\{W\}over a vocabularyWW;ρ\\rhoholds routing fields \(KPI/playbook/schema\-mapping reference stems, a persona example, core references, and aforce\_dsflag\);YYare standing\-report templates;LLare deterministic library modules; and𝑐𝑓𝑔\\mathit\{cfg\}holds text budgets\. The*registry*K=\{k:SKILL\.mdpresent\}K=\\\{k:\\text\{\{\\small SKILL\.md\} present\}\\\}is a pure function of disk\.

### 3\.2Selection: deterministic and fail\-open

Writea⊑Ba\\sqsubseteq Bfor “aaoccurs as a lowercased substring in blobBB\.” Define, for a skillkkand schemaΣ\\Sigma, the hit counts

t=\|\{a∈any\_table:a⊑T\}\|,c=\|\{a∈any\_column:a⊑𝐶𝑜𝑙\}\|,t=\|\\\{a\\in\\textsf\{any\\\_table\}:a\\sqsubseteq T\\\}\|,\\quad c=\|\\\{a\\in\\textsf\{any\\\_column\}:a\\sqsubseteq\\mathit\{Col\}\\\}\|,v=\|\{a∈any\_value:a⊑𝑉𝑎𝑙\}\|,w=\|\{a∈keywords:a⊑\(T∪𝐶𝑜𝑙\)\}\|,v=\|\\\{a\\in\\textsf\{any\\\_value\}:a\\sqsubseteq\\mathit\{Val\}\\\}\|,\\quad w=\|\\\{a\\in\\textsf\{keywords\}:a\\sqsubseteq\(T\\cup\\mathit\{Col\}\)\\\}\|,and the score

score​\(k,Σ\)=2​t\+c\+2​v\+w\.\\mathrm\{score\}\(k,\\Sigma\)=2t\+c\+2v\+w\.\(1\)A skill*qualifies*iff

\(t≥1∨v≥1\)∧score​\(k,Σ\)≥min\_score​\(k\)\.\\big\(t\\geq 1\\;\\lor\\;v\\geq 1\\big\)\\;\\land\\;\\mathrm\{score\}\(k,\\Sigma\)\\geq\\textsf\{min\\\_score\}\(k\)\.\(2\)The table\-or\-value requirement in[Eq\.˜2](https://arxiv.org/html/2608.28594#S3.E2)is what prevents a lone coincidental column \(a strayvendorcolumn on an unrelated finance schema\) from auto\-triggering an expert; the*value*disjunct is what lets a distinctive cell string \(a Tally accounting group such as “Sundry Debtors”\) stand in for a table\-name hit on schemas whose*names*are opaque\. In the live pathσ\\sigmais evaluated with𝑉𝑎𝑙=∅\\mathit\{Val\}=\\varnothing\(the chat selector scores over names only\); the*value*disjunct is exercised by the offline detector over sampled cells \([Section˜5](https://arxiv.org/html/2608.28594#S5)\) and folded back into the live assignment by the onboarding chain’s persistence edge \([Section˜7](https://arxiv.org/html/2608.28594#S7)\)\. The active skill is

σ​\(c,d\)=\{assign​\(c,d\)explicit assignment exists \(⊥if disabled\)arg​mink​qualifying⁡\(−score​\(k,Σ\),−min\_score​\(k\),𝑠𝑙𝑢𝑔​\(k\)\)else, if any qualifies⊥otherwise,\\sigma\(c,d\)=\\begin\{cases\}\\mathrm\{assign\}\(c,d\)&\\text\{explicit assignment exists \(\}\\bot\\text\{ if disabled\)\}\\\\\[2\.0pt\] \\displaystyle\\operatorname\*\{arg\\,min\}\_\{k\\ \\text\{qualifying\}\}\\big\(\-\\mathrm\{score\}\(k,\\Sigma\),\\,\-\\textsf\{min\\\_score\}\(k\),\\,\\mathit\{slug\}\(k\)\\big\)&\\text\{else, if any qualifies\}\\\\\[6\.0pt\] \\bot&\\text\{otherwise,\}\\end\{cases\}\(3\)where thearg​min\\operatorname\*\{arg\\,min\}is over the lexicographic order shown: highest score, then the skill that cleared a stricter bar \(highermin\_score, the stronger claim relative to its own threshold\), then the alphabetically\-first slug, making selection deterministic and independent of registry order\.σ\\sigmais gated by a global flag, memoized with a short positive and a deliberately shorter negative cache window, and yields⊥\\boton*any*error; it is fail\-open by construction\. The resolved slug is written once intox\.active\_skillx\.\\textsf\{active\\\_skill\}in stages1s\_\{1\}and carried as a lightweight string thereafter\.

### 3\.3The injection operator and its no\-op property

For a stagesis\_\{i\}with active skillk=σ​\(c,d\)k=\\sigma\(c,d\), define the injected prompt

inji​\(𝑝𝑟𝑜𝑚𝑝𝑡,x\)=𝑝𝑟𝑜𝑚𝑝𝑡⊕renderi​\(k,x\),\\mathrm\{inj\}\_\{i\}\(\\mathit\{prompt\},x\)=\\mathit\{prompt\}\\;\\oplus\\;\\mathrm\{render\}\_\{i\}\(k,x\),\(4\)where⊕\\oplussplices facet/reference text at the stage’s anchor \(for the coder,*before*any custom\-prompt override marker\) andrenderi​\(⊥,x\)=ε\\mathrm\{render\}\_\{i\}\(\\bot,x\)=\\varepsilon\. As an example, the coder render composes the methodology and chart facets, a keyword\-selected reference, and a shared reasoning\-discipline block:

rendercoder=H⊕F​\(coder\)⊕F​\(chart\)⊕refsel​\(k,q\)⊕DISC,\\mathrm\{render\}\_\{\\textsf\{coder\}\}=H\\oplus F\(\\textsf\{coder\}\)\\oplus F\(\\textsf\{chart\}\)\\oplus\\mathrm\{refsel\}\(k,q\)\\oplus\\mathrm\{DISC\},withrefsel​\(k,q\)=arg​maxstem⁡\|\{w∈κ​\(stem\):w⊑q\}\|\\mathrm\{refsel\}\(k,q\)=\\operatorname\*\{arg\\,max\}\_\{\\textit\{stem\}\}\|\\\{w\\in\\kappa\(\\textit\{stem\}\):w\\sqsubseteq q\\\}\|over the \(enhanced\) questionqq, truncated to a chat budget; the offline report path uses the*same*routingκ\\kappabut admits multiple*full*references under a larger budget\.

###### Proposition 1\(No\-op additivity / backward compatibility\)\.

Assume each stage’s dependence on the active skill is confined toinji\\mathrm\{inj\}\_\{i\}\(the slug enterssis\_\{i\}only through the injected prompt\)\. Ifσ​\(c,d\)=⊥\\sigma\(c,d\)=\\botthen for every stageiiand every𝑝𝑟𝑜𝑚𝑝𝑡\\mathit\{prompt\},inji​\(𝑝𝑟𝑜𝑚𝑝𝑡,x\)=𝑝𝑟𝑜𝑚𝑝𝑡\\mathrm\{inj\}\_\{i\}\(\\mathit\{prompt\},x\)=\\mathit\{prompt\}, and the skill\-augmented pipeline equals the baseline pipeline on every executed sub\-sequence ofPP\.

###### Proof\.

renderi​\(⊥,x\)=ε\\mathrm\{render\}\_\{i\}\(\\bot,x\)=\\varepsilonfor alliiby definition andp⊕ε=pp\\oplus\\varepsilon=p, so each stage receives an unchanged prompt; by the factoring hypothesis the slug affectssis\_\{i\}only through that prompt, so eachsis\_\{i\}is unchanged as a function\. Composing unchanged stages along any executed route leaves the pipeline unchanged\. ∎

[Proposition˜1](https://arxiv.org/html/2608.28594#Thmproposition1)is the property that makes the entire subsystem safe to ship: with no skill active, or any failure in resolution, the product is byte\-identical to its skill\-free baseline\.

### 3\.4Offline knowledge compilation

LetBBbe a read\-only DuckDB\-on\-parquet backend overΣ\\Sigma\. For each tableτ\\tau, a probe agentAAruns a self\-terminating loop that emits actions from a tool setToolscomprisingget\_schema,profile\_columns,sample\_values,run\_sql,record\_finding, andsubmit; a deterministic coverage criticγcov\\gamma\_\{\\mathrm\{cov\}\}maps a submission to\{ok,gaps\}\\\{\\textsf\{ok\},\\textsf\{gaps\}\\\}, and an*optional*LLM depth criticγdep\\gamma\_\{\\mathrm\{dep\}\}\(fail\-open, returningdeepon any error\) may demand a deeper pass;converged​\(τ\)\\mathrm\{converged\}\(\\tau\)holds whenγcov\\gamma\_\{\\mathrm\{cov\}\}is satisfied and, where enabled,γdep\\gamma\_\{\\mathrm\{dep\}\}is too\. The compiler iterates over the set of unconverged tables:

W0=T,Wr\+1=\{τ∈Wr:¬converged​\(probe​\(τ,𝑓𝑒𝑒𝑑𝑏𝑎𝑐𝑘r​\(τ\)\)\)\},r=1,…,R,W\_\{0\}=T,\\qquad W\_\{r\+1\}=\\\{\\tau\\in W\_\{r\}:\\lnot\\,\\mathrm\{converged\}\(\\mathrm\{probe\}\(\\tau,\\mathit\{feedback\}\_\{r\}\(\\tau\)\)\)\\\},\\quad r=1,\\dots,R,\(5\)with a round capRR\(and a wall\-clock backstop\), where a*keep\-best*policy guarantees a table never regresses once converged\.

###### Proposition 2\(Termination\)\.

The sequence\(Wr\)\(W\_\{r\}\)is monotone decreasing \(Wr\+1⊆WrW\_\{r\+1\}\\subseteq W\_\{r\}\) and the loop terminates in at mostRRrounds\. The compiled depth isdeepiffWR=∅W\_\{R\}=\\emptyset, elsepartial\.

###### Proof\.

By[Eq\.˜5](https://arxiv.org/html/2608.28594#S3.E5),Wr\+1W\_\{r\+1\}is a subset ofWrW\_\{r\}by construction; under keep\-best, a table that converged in roundrris excluded from all laterWr′W\_\{r^\{\\prime\}\},r′\>rr^\{\\prime\}\>r, so no element re\-enters\. A non\-increasing sequence of finite sets bounded below by∅\\emptyset, evaluated for at mostRRrounds, terminates\. The depth label reads offWRW\_\{R\}\. ∎

#### Join validation\.

For a candidate joinτa\.x→τb\.y\\tau\_\{a\}\.x\\to\\tau\_\{b\}\.yproposed by any of four signals \(same name, shared skill\-entity binding, data\-driven identifier discovery, or an LLM proposal told that names may differ\), accept iff

cover​\(x→y\)=\|distinct​\(x\)∩distinct​\(y\)\|\|distinct​\(x\)\|≥0\.5,\\mathrm\{cover\}\(x\\\!\\to\\\!y\)=\\frac\{\|\\mathrm\{distinct\}\(x\)\\cap\\mathrm\{distinct\}\(y\)\|\}\{\|\\mathrm\{distinct\}\(x\)\|\}\\geq 0\.5,\(6\)the column types are compatible, andτb\.y\\tau\_\{b\}\.yis either data\-measureduniqueor a single\-column primary key \(a member of a*composite*key is rejected, being non\-unique alone\)\. Coverage is directional, the fraction of child\-key values with a referent in the candidate parent, so the0\.50\.5floor tolerates the partial referential integrity common in raw enterprise extracts, while the uniqueness requirement onτb\.y\\tau\_\{b\}\.yfixes the join’s direction\. Value overlap establishes the*precision*of accepted joins; recall is bounded by the four candidate generators, and we do not claim a complete relationship graph\. Each accepted relationship carries its coverage, inferred cardinality, a confidence, and anevidence sqlwitness\. The output𝒦​\(c,d\)=\(per\-table knowledge,validated joins,glossary\)\\mathcal\{K\}\(c,d\)=\(\\text\{per\-table knowledge\},\\text\{validated joins\},\\text\{glossary\}\)is persisted durably; a compact briefbrief​\(𝒦\)\\mathrm\{brief\}\(\\mathcal\{K\}\)is shared by the report planner and the question generator\.

### 3\.5Admissibility: the trust contract

A report viewvvdefines metrics\{\(ℓ,ref=\(step,col,agg\),evidence\_sql\)\}\\\{\(\\ell,\\mathrm\{ref\}=\(\\textit\{step\},\\textit\{col\},\\textit\{agg\}\),\\textsc\{evidence\\\_sql\}\)\\\}\. Let𝑑𝑒𝑟𝑖𝑣𝑒𝑑=agg\(step\.col\)\\mathit\{derived\}=\\textit\{agg\}\(\\textit\{step\}\.\\textit\{col\}\)be the value the view computed and𝑒𝑣=scalar​\(run​\_​sql​\(evidence\_sql\)\)\\mathit\{ev\}=\\mathrm\{scalar\}\(\\mathrm\{run\\\_sql\}\(\\textsc\{evidence\\\_sql\}\)\)the value obtained by independently re\-executing the evidence query\.

###### Definition 2\(Admissible view\)\.

vvis*admissible*iff for every metric

\|𝑑𝑒𝑟𝑖𝑣𝑒𝑑−𝑒𝑣\|max⁡\(\|𝑑𝑒𝑟𝑖𝑣𝑒𝑑\|,\|𝑒𝑣\|,ϵ\)≤0\.01,\\frac\{\|\\mathit\{derived\}\-\\mathit\{ev\}\|\}\{\\max\(\|\\mathit\{derived\}\|,\|\\mathit\{ev\}\|,\\epsilon\)\}\\leq 0\.01,\(7\)and every chart/table reference resolves to a defined step and a real column\.

A tolerance is needed because𝑑𝑒𝑟𝑖𝑣𝑒𝑑\\mathit\{derived\}and𝑒𝑣\\mathit\{ev\}follow*different*execution paths \(recipe\-step aggregation versus a standalone scalar query\) and can differ by float accumulation and rounding; the1%1\\%relative bound absorbs this while still catching the order\-of\-magnitude and wrong\-column errors that matter\.111The1%1\\%bound is a configurable default and a small positiveϵ\\epsilonfloors the denominator against division by zero; the guarantee boundary is thus a parameter, not a universal constant\.A non\-admissible view is*suppressed*; a report whose views are all non\-admissible is*deferred*\(hidden\)\. The published artifact is therefore a*durable recipe*r​\(v\)r\(v\)\(a SQL\-step DAG, metric definitions, deterministic chart specifications, and filters\), and refresh is the*zero\-LLM, zero\-kernel*evaluationexec​\(r,𝑝𝑎𝑟𝑎𝑚𝑠\)\\mathrm\{exec\}\(r,\\mathit\{params\}\); re\-narration fires only whenfingerprint​\(exec​\(r,∅\)\)\\mathrm\{fingerprint\}\(\\mathrm\{exec\}\(r,\\varnothing\)\)changes\.

#### The proactive loop\.

Onboarding induces the chain

upload→fast​\(Σ\)→𝒦=compile​\(Σ,σ\)→persist​σ\\displaystyle\\textsf\{upload\}\\to\\textsf\{fast\}\(\\Sigma\)\\to\\mathcal\{K\}=\\mathrm\{compile\}\(\\Sigma,\\sigma\)\\to\\textsf\{persist\}\\,\\sigma→Reports=\{v​admissible:v∈Y​\(σ\)​over​𝒦\}\\displaystyle\\quad\\to\\textsf\{Reports\}=\\\{v\\ \\text\{admissible\}:v\\in Y\(\\sigma\)\\ \\text\{over\}\\ \\mathcal\{K\}\\\}→Questions=gen​\(σ,brief​\(𝒦\),Reports\),\\displaystyle\\quad\\to\\textsf\{Questions\}=\\mathrm\{gen\}\(\\sigma,\\mathrm\{brief\}\(\\mathcal\{K\}\),\\textsf\{Reports\}\),and a user click on a published numbernnin viewvvmaps to a natural\-language queryq​\(n,v\)q\(n,v\)that re\-entersPP, grounded in the same𝒦\\mathcal\{K\}\(whose validated joins are authoritative incollate\)\. The system is proactive in that it emits\(Reports,Questions\)\(\\textsf\{Reports\},\\textsf\{Questions\}\)*before*anyqq, and reactive deep dives areq↦P​\(q\|𝒦,σ\)q\\mapsto P\(q\\,\|\\,\\mathcal\{K\},\\sigma\)\.

## 4The Skill Abstraction

A skill is the unit that makes domain expertise portable\. It is a folder \(not a prompt, not a fine\-tune, not a database row\), so it is version\-controlled, reviewable, and shippable like code, yet it influences every reasoning step in the system\.[Figure˜1](https://arxiv.org/html/2608.28594#S4.F1)shows its anatomy\.

skills/ca\_expert/SKILL\.md: manifestmatch,min\_score=5,force\_ds, reference fields, personafacets/: per\-stage prompt fragmentsrouter · scout · coder · chart · narrator · report\_planner · report\_analystreferences/: deep knowledgegst · direct\_tax · statements&ratios · tally\_data\_model · playbook · auditreports\.yaml: 7 standing report templateslib/accounting\.py: deterministic compute \(opt\-in\)router · scoutcoder · chartnarratordeep explorer \(skill lens\)report planner · analystquestion generatorchat injects*one*short reference excerpt; reports inject*multiple*full references; a shared reasoning\-discipline block attaches to every reasoning lensFigure 1:Anatomy of a domain\-expert skill pack \(the chartered\-accountant skill\) and where each part is injected\. The same folder feeds the live pipeline, the offline explorer, both report engines, and the question generator\.### 4\.1Anatomy of a skill pack

A pack contains aSKILL\.mdmanifest \(YAML frontmatter plus a body\), afacets/directory of per\-stage prompt fragments, areferences/directory of deep\-knowledge documents, areports\.yamlof standing\-report templates, and an optionallib/of deterministic Python\. The manifest declares only*lightweight routing configuration*: the auto\-detect patterns, a keyword\-to\-reference map, which references are “core” \(always loaded for reports\), which reference supplies the KPI vocabulary and which the review cadence, a persona example for narration, and a flag that routes the coder to the data\-science agent so the skill’s modeling methods are available\. The heavy text \(facets, references\) is loaded into an in\-memory object; the run context carries only the slug\. We ship three packs \(a Chartered Accountant expert, a Chief Marketing Officer expert, and a Supply\-Chain expert\) to demonstrate that the framework is not hardcoded to one domain; the structure, not the result, is what the second and third packs evidence\.

Listing 1:The frontmatter of the chartered\-accountant skill \(excerpt\)\. Routing fields andmatchpatterns are declarative; the value tokens enable detection on opaque schemas\.name:ca\_expert

force\_ds\_agent:true

persona\_example:businessownerorCFO

kpi\_reference:financial\_statements\_and\_ratios

playbook\_reference:ca\_practice\_playbook

schema\_mapping\_reference:tally\_data\_model\#deep\-explorerrecognitionlens

reference\_keywords:

indian\_gst:\[gst,gstr,itc,cgst,sgst,igst,hsn,"placeofsupply","3b"\]

indian\_direct\_tax:\[tds,tcs,"incometax","advancetax","44ab","194",\.\.\.\]

financial\_statements\_and\_ratios:\[ratio,"balancesheet",roce,"scheduleiii",\.\.\.\]

match:

min\_score:5

any\_table:\[ledger,voucher,daybook,trial\_balance,gstr,"\_tds\_",tally,\.\.\.\]

any\_column:\[ledger,voucher\_type,voucher\_no,debit,credit,\.\.\.\]

any\_value:\["sundrydebtors","sundrycreditors","duties&taxes",\.\.\.\]\#opaque\-schemafallback

### 4\.2Registry and selection

The registry is a lazy, thread\-locked process singleton built once by scanning theskills/directory; it has no database dependency, so it is safe to import from any node, including the headless report jobs, and multi\-worker deployments each build an identical registry from the same files\. A malformed pack is logged and skipped rather than breaking the rest\.

Selection follows[Eq\.˜3](https://arxiv.org/html/2608.28594#S3.E3)\. An explicit assignment \(an admin override, or a previously cached auto\-detection\) wins first; otherwise the system loads table and column names, preferring a local parquet schema cache and falling back to a unified cross\-connector schema config, so detection is source\-agnostic across uploaded files, SAP, Snowflake, Postgres, and others, and runs the scorer of[Eqs\.˜1](https://arxiv.org/html/2608.28594#S3.E1)and[2](https://arxiv.org/html/2608.28594#S3.E2)\. On a confident hit the result is persisted back as anauto\_detectedassignment so subsequent lookups are a single fast read and an operator can see and override the choice\.[Algorithm˜1](https://arxiv.org/html/2608.28594#algorithm1)states the resolution\.

Input:client

cc, dataset

dd; registry

KK
Output:active slug or

⊥\\bot
1if*notshould\_use\_skills\(c\)\(c\)*then

2return

⊥\\bot
3if*cache hit for\(c,d\)\(c,d\)*then

4returncached value

5if*explicit assignmentAAexists*then

6if*AAdisabled*thenreturn

⊥\\bot
7if*A\.𝑠𝑙𝑢𝑔A\.\\mathit\{slug\}valid*thenreturn

A\.𝑠𝑙𝑢𝑔A\.\\mathit\{slug\}
8

9

\(T,𝐶𝑜𝑙\)←\(T,\\mathit\{Col\}\)\\leftarrowschema names \(parquet cache, else connector config\)

m←match\_schema​\(T,𝐶𝑜𝑙,𝑉𝑎𝑙\)m\\leftarrow\\texttt\{\\small match\\\_schema\}\(T,\\mathit\{Col\},\\mathit\{Val\}\)

//[Eqs\.˜1](https://arxiv.org/html/2608.28594#S3.E1),[2](https://arxiv.org/html/2608.28594#S3.E2)and[3](https://arxiv.org/html/2608.28594#S3.E3)

10if*m≠⊥m\\neq\\bot*then

11persist

mmasauto\_detected;return

m\.𝑠𝑙𝑢𝑔m\.\\mathit\{slug\}
return

⊥\\bot
//any exception in the above also returns⊥\\bot

Algorithm 1Deterministic, fail\-open skill selectionσ​\(c,d\)\\sigma\(c,d\)
### 4\.3Cross\-cutting injection

[Equation˜4](https://arxiv.org/html/2608.28594#S3.E4)is realized by one smallrender\_\*helper per node, each a no\-op when no skill is active \([Proposition˜1](https://arxiv.org/html/2608.28594#Thmproposition1)\)\. The router receives a domain lens for question rewriting and table routing; each per\-table scout prompt is annotated with the columns the expert cares about; the coder system prompt is prefixed, ahead of any client custom prompt, with a methodology block that bundles the coder and chart facets, plus a single keyword\-selected reference excerpt; and the narrator is framed to brief the skill’s persona\. A subtlety that matters in practice: the*chat*path splices exactly one reference excerpt under a tight budget \(latency is bounded\), whereas the offline report path uses the*same*keyword routing but loads multiple*full*references under a far larger budget, because reports are correctness\-first batch jobs that can afford the context\. Finally, a shared*reasoning\-discipline*block, demanding explicit confidence levels, named blind spots, confounders\-before\-conclusions, and a falsification test for each headline claim, is inherited by every skill at every reasoning lens, attaching an explicit confidence/blind\-spot/falsification discipline to each piece of generated reasoning\. An optional, off\-by\-default hook can additionally inject a skill’s deterministic Python into the kernel bootstrap, wrapped so that a broken library is non\-fatal and the live product is unchanged until an operator opts in\.

### 4\.4Toward a marketplace of expert skills

The abstraction is deliberately open\-ended, and three properties make the catalogue*extensible by addition*rather than by engineering\. \(i\) A skill is a self\-contained folder with no database dependency, so authoring a new expert is a content task \(writing the manifest, facets, references, and report templates\), not a code change; the registry discovers it on a directory scan\. \(ii\) Selection is a deterministicNN\-way score with a stable tiebreak \([Eq\.˜3](https://arxiv.org/html/2608.28594#S3.E3)\), so adding the tenth or hundredth skill does not perturb the resolution of the first; the catalogue scales without a combinatorial routing problem\. \(iii\) The no\-op property \([Proposition˜1](https://arxiv.org/html/2608.28594#Thmproposition1)\) guarantees that an unmatched or malformed pack cannot degrade the base product, so packs can be published, versioned, and retired independently and safely\. Together these turn the framework into a substrate for a*marketplace of domain experts*: each pack encodes the playbook of a senior practitioner \(a PhD\-level or veteran industry specialist\) and the three we ship \(a Chartered Accountant, a Chief Marketing Officer, and a Supply\-Chain expert\) are seeds of a catalogue that could span equity research, actuarial and clinical analysis, FP&A, pricing and revenue management, risk and audit, and beyond\. A tenant connects data, the relevant expert is selected automatically, and the same machinery that serves one expert then serves an arbitrary library of them\. We make no claim to have populated such a marketplace; the present catalogue is small and human\-authored \([Section˜11](https://arxiv.org/html/2608.28594#S11)\)\. We claim only that the architecture is built to host one\.

## 5Persistent Agentic Schema Exploration

Before the system can apply domain expertise, it must build structural knowledge of the data, profiling each column, testing encoding hypotheses, and validating joins\. A fast one\-shot pass writes a one\-line description per table at onboarding; the knowledge\-compilation phase is a second, slower pass that actually digs in\. Its design is borrowed directly from how a coding agent explores a codebase: give a model a small set of tools, let it choose actions, feed back observations, and let an*independent*check decide when it is done\[[59](https://arxiv.org/html/2608.28594#bib.bib11),[47](https://arxiv.org/html/2608.28594#bib.bib12)\]\.

probe agentJSON action over DuckDBsubmitterminal actioncoveragecritic\(code\)assemble \+upsert to storetool resultokgaps: “you missed column X”anti\-runaway guards: doom\-loop \(3×3\\timessame action\) · diminishing\-returns nudge · step ceiling · history trimW1W\_\{1\}:\|T\|\|T\|tablesW2⊆W1W\_\{2\}\\subseteq W\_\{1\}W3⊆W2W\_\{3\}\\subseteq W\_\{2\}deepiffWR=∅W\_\{R\}=\\emptysetelsepartialroundroundself\-healing convergence \([Proposition˜2](https://arxiv.org/html/2608.28594#Thmproposition2)\),R≤3R\\leq 34 candidate signalssame\-name · skill\-entity · data\-driven id · LLM \(“names may differ”\)overlap≥0\.5\\geq 0\.5type\-compat\.parent uniqueRelationship\{overlap, cardinality, confidence, evidence\_sql\}DuckDB value\-overlapacceptFigure 2:Knowledge compilation\. Top: the per\-table probe loop, where “continue” is the default and “stop” is an explicit terminal action gated by a*deterministic*coverage critic, bounded by anti\-runaway guards\. Middle: the self\-healing convergence loop, whose pending set provably shrinks\. Bottom: join discovery, where every candidate is*proven*on the data by value overlap\.### 5\.1The per\-table probe loop

One agent runs per table \(fanned out with bounded concurrency\)\. It is handed the table’s columns and types, the tool contract, and, if a skill matched, an expert “lens” describing what matters in this domain\. It replies with a single action in JSON \(the providers do not reliably support native tool calling, so actions are dispatched from JSON, the same mechanism the live coder uses\); the system executes that action against DuckDB over the parquet files \(*zero*load on the client’s production database\) and returns the result as the next observation\. The agent profiles columns, samples values, tests encoding hypotheses \(“is this period field reallyYYYYMMtext?”\), and checks keys, until it calls the terminalsubmitaction with a full characterization\.

#### Stop is gated, not self\-declared\.

Crucially, the decision to stop is not the model’s to make alone\. A*deterministic*coverage critic \(plain code, not an LLM, so it is fast, free, and never hallucinates\) checks whether every column was profiled and typed, whether a primary key and grain were identified, and whether the skill’s expected entities were found\. If something is missing it returns a specific gap \(“you missed columnXX”\) and the loop continues; only when the coverage critic is satisfied is the table assembled and upserted\. An*optional*LLM “depth” critic may then require a second, deeper pass; it is fail\-open \(treating any error as satisfied\), so it can raise quality but never blocks termination \([Proposition˜2](https://arxiv.org/html/2608.28594#Thmproposition2)\)\. This mirrors the “verification gate before done” pattern and directly addresses the finding that intrinsic self\-assessment is unreliable\[[25](https://arxiv.org/html/2608.28594#bib.bib49)\]: the in\-loop gate is external and deterministic\. Anti\-runaway guards \(breaking on the same action repeated three times, a diminishing\-returns nudge, a per\-table step ceiling, and trimming of old observations so context stays small\) bound the loop, and a keep\-best policy ensures the self\-healing re\-attempts of[Eq\.˜5](https://arxiv.org/html/2608.28594#S3.E5)never regress a converged table\. The honest depth label follows[Proposition˜2](https://arxiv.org/html/2608.28594#Thmproposition2): a run isdeeponly if*every*table converged, elsepartial\.

### 5\.2Joins proven on the data

Cross\-table joins are where naive schema reasoning fails, because matching columns frequently have*different names*across tables\. The explorer generates candidates from four generic signals \(identical names, a shared skill\-entity binding, data\-driven discovery of identifier\-like columns, and an LLM proposal explicitly told that names may differ and to match by meaning and overlapping sample values\) and then*proves*each one on the data with a DuckDB value\-overlap query \([Eq\.˜6](https://arxiv.org/html/2608.28594#S3.E6)\), keeping only joins whose values actually overlap, whose types are compatible, and whose parent side has*measured*uniqueness\. The use of measured cardinality rather than a declared primary\-key flag is deliberate: a member of a composite key is not unique on its own and must not be treated as a parent\. Each surviving relationship carries its overlap percentage, cardinality, a confidence, and an evidence query, and these validated joins are later treated as authoritative by the live pipeline’scollatestage\.

### 5\.3Persistence and economics

Findings are written to a per\-\(client, dataset\) document in the operational store the moment each table is done \(durable and resumable\) and exported to a JSON file for backup\. Human curation \(notes on a database, table, or column, and manually edited relationships\) is preserved across re\-runs, and refreshes are capped \(a small fixed number\) to bound cost; the live pipeline reads the knowledge through a short in\-process cache\. The detected skill is persisted back to the assignment store so the live chat selector \(which keys off names\) and the value\-based detector used here agree, an edge in the onboarding chain we revisit in[Section˜7](https://arxiv.org/html/2608.28594#S7)\. The cost of compilation scales with the number of tables and their width,*not*with row count, because the data probing is DuckDB work and only the agent’s reasoning consumes tokens; it is a one\-time, amortized cost rather than a per\-query one\.

## 6Execution\-Verified Expert Report Synthesis

Standing expert reports are the proactive product’s centerpiece: pre\-computed, domain\-expert views a user reads*first*\. Because they are shown without a prompting question, correctness matters more than latency, and the engine is modeled on the explorer \(a planned, critic\-gated agent\) rather than on the live coder\.[Figure˜3](https://arxiv.org/html/2608.28594#S6.F3)shows the pipeline;LABEL:lst:reportshows a template view; the trust core is[Definition˜2](https://arxiv.org/html/2608.28594#Thmdefinition2)\.

Lead\-analystplannertemplate = hypothesis \+ quality floorSection analystsparallel; define SQL steps \+ KPIs \+ evidence\_sqlCritic:re\-runevidence\_sqlassemble recipe→\\tonarrate→\\topublishsuppressview33–66cutspass \(≤1%\\leq 1\\%\)blockreport with no admissible view⇒\\Rightarrowdefer\(hidden\)Durable recipeSQL DAG · KPI defs · chart specs · filtersrefresh · filter ·chart\-click drill\-downzero\-LLM / zero\-kernelFigure 3:Execution\-verified report synthesis\. Every metric carries an evidence query that an independent, non\-LLM critic re\-executes; a view that fails[Definition˜2](https://arxiv.org/html/2608.28594#Thmdefinition2)is suppressed, and a report with no admissible view is deferred\. The published artifact is a durable recipe that refreshes and drills down at zero LLM and zero kernel cost\.### 6\.1Plan, fan out, verify, narrate

An adaptive lead\-analyst planner treats areports\.yamltemplate as a*hypothesis and a quality floor*, not a script: it probes the data and expands the report into a handful of verifiable cuts that the actual columns support\. A section analyst is then dispatched per view \(bounded concurrency\), each defining durable SQL steps and emitting metrics with an evidence query attached\. The authoritative step is the critic: for each metric it*independently*re\-executes the evidence SQL and tolerance\-matches the result against the value the view derived \([Eq\.˜7](https://arxiv.org/html/2608.28594#S3.E7)\); a value/evidence mismatch, a broken reference, or an evidence\-query error is a*block*\(a non\-numeric result is a block for a metric that declares a numeric unit\)\. This is execution\-grounded verification in the spirit of CRITIC and self\-debugging\[[18](https://arxiv.org/html/2608.28594#bib.bib50),[8](https://arxiv.org/html/2608.28594#bib.bib51)\], applied not to fix code but to*gate publication*: a number that does not reproduce against its own evidence query is never shown\.

#### The failure mode is silence, not error\.

A view that fails verification is suppressed; a report with no surviving view is deferred and simply not displayed\. The system never surfaces a raw error to a user, and, within the boundary stated below, never publishes a number it could not reproduce\. The honest claim is not “zero hallucination” or “always correct”; it is*suppress\-or\-defer*: an unverifiable figure is withheld\. And the boundary is precise:[Eq\.˜7](https://arxiv.org/html/2608.28594#S3.E7)certifies that a published number*matches its declared evidence SQL*, not that the metric’s business definition is the semantically right one\. Verification guards arithmetic faithfulness, not definitional correctness\.

Listing 2:A standing\-report template view \(the GST summary, excerpt\)\. The template is client\-agnostic instructions; the engine maps it onto whatever books the client has, then verifies and saves a re\-runnable recipe\.\-id:gst\_summary\_reconciliation

category:GST

applies\_when:DatasethasGSTfields\(CGST/SGST/IGST,taxablevalue\)\.\.\.

kpi\_hints:\[output\_tax,input\_tax,net\_gst\_payable,itc\_available\]

views:

\-id:overview

prompt:\>

AsaGSTexpert,summarisetheGSTpositionforthetaxperiod\.Mapthe

schematotaxablevalue,CGST/SGST/IGSTandsales\-vs\-purchasedirection\.

Reportoutputtax,inputtax/ITCandNETGSTpayable,splitbyrate\.Where

aportalfigure\(GSTR\-2B\)isabsent,DECLAREtheblindspotratherthan

assumeallinputtaxiscreditable\.\.\.\.

sections:\[kpi\_strip,chart,table,insight\]

### 6\.2The durable recipe

Publication produces a*recipe*: a SQL\-step DAG, metric definitions \(each a reference into a step plus an aggregation and its evidence query\), deterministic chart specifications, a drill\-down table, and filters\. The recipe, not an LLM, is what refresh re\-runs: refreshing, filtering, and chart\-click drill\-down all evaluate the recipe with zero LLM and zero kernel cost, and because charts, KPI strip, and table all read from the*same*step rows, they cannot disagree\. Re\-narration is gated by a data fingerprint, so the language model is invoked again only when the underlying numbers actually change\. The published document is shaped to match the platform’s existing report contract exactly, carrying the recipe, a “how we computed this” provenance view that exposes the evidence SQL, and a per\-view confidence tier additively, so the engine is a drop\-in behind a flag, fully backward\-compatible with the legacy generator\.

## 7Proactive Question Generation and the Loop

The final step is to tell the user*what to ask*\. Suggested questions are generated from two fused sources: the matched skill’s report themes, KPI catalogue, and review\-cadence playbook \(the expert’s agenda\), and the*same*validated\-knowledge brief the report planner consumes \(so the questions are answerable against real entities, grains, and joins\)\. The generator builds roughly thirty diverse questions, then rewrites them into standalone analyst phrasing using business vocabulary \(never raw column names\) and writes them to a cache\. The prompt targets the review\-cadence questions a senior domain leader would ask to run the business, scoped to the entities the schema exposes; we make no claim about how end users perceive them\. Everything is fail\-open: missing metadata, skill, or knowledge degrades to schema\-only questions, and finally to a fixed set of generic fallbacks, so the home page is never empty\.

Data connected /uploadedFast \+ Deep explorationvalidated knowledge𝒦\\mathcal\{K\}, skill auto\-persistedStanding expert reportsverified KPIsSuggested questionsmirror report KPIsUser reads numbers,clicks a figureLive pipelinePPverified deep divee1fast→\\todeepe2persist skille4on successcomposed NL promptq\|𝒦,σq\\,\|\\,\\mathcal\{K\},\\sigmae3deep→\\toreports \(only on convergence\)No prior user query is required: the loop runs to completion before the blank query box is ever used\.Figure 4:The proactive\-analyst loop and the fully\-automatic onboarding chain\. Four trigger edges \(e1–e4\) wire onboarding end\-to\-end; a clicked number is handed back to the live pipeline as an on\-topic deep\-dive query grounded in the same compiled knowledge\.### 7\.1The fully\-automatic onboarding chain

What makes this a*loop*rather than a set of features is that the edges fire automatically \([Fig\.˜4](https://arxiv.org/html/2608.28594#S7.F4)\)\. Connecting or uploading data triggers the fast pass;e1the fast pass triggers deep exploration;e2deep exploration persists the detected skill into the assignment store \(bridging the name\-based chat selector and the value\-based deep detector so both agree\);e3deep exploration triggers report generation*only after*convergence; ande4successful report generation triggers question regeneration so the questions mirror the freshly verified report KPIs\. Each edge is fire\-and\-forget and fail\-open, and autonomy is human\-overridable throughout: an administrator can assign or disable a skill, edit knowledge notes, and trigger a capped refresh\.

### 7\.2Closing the loop in the interface

The user lands not on a blank box but on a board of finished analyses: an operational brief, a hero summary, alert cards, each with a small visualization rendered from the*real*published figure\. Reading a number raises the natural follow\-up “why?”, and there are two paths\. The first is an in\-page, zero\-LLM drill: clicking a chart re\-evaluates the recipe at a finer grain\. The second hands control back to the live pipeline: a composed natural\-language prompt is marshaled into the streaming chat, where it is answered against the same validated knowledge, closing the loop from*reports*to*numbers*to*questions*to*verified deep dive*\. Because the suggestions come from a domain\-expert prior that exists on day one and are generated against the auto\-explored schema, they sidestep both failure modes of prior art: they need no query logs\[[6](https://arxiv.org/html/2608.28594#bib.bib7),[54](https://arxiv.org/html/2608.28594#bib.bib8)\]and they are role\-relevant rather than merely statistically salient\[[15](https://arxiv.org/html/2608.28594#bib.bib9)\]\.

## 8System and Implementation

We describe the system at the level of its*principles and contracts*\(the abstractions, the selection and admissibility rules, and the engineering posture that make the architecture work\) rather than as a turnkey blueprint; internal prompt text, tuning constants, and module\-level wiring are elided as implementation detail\. With that scope set, the live regime is a custom, durable, multi\-stage streaming pipeline \(the stages of[Section˜3](https://arxiv.org/html/2608.28594#S3)\)\. A run is a detached, durable task that writes its events to a per\-run append\-only log; the HTTP response body is a replay\-then\-tail reader over that log, which makes reconnection and resume natural and decouples the long\-running analysis from event delivery\. The active skill is resolved exactly once, inguard, in parallel with relevance guards, and threaded downstream as a lightweight slug\. Multi\-tenancy is strict: every datastore query, storage path, and job is scoped by a client identifier with no defaults, table\-level access control is enforced at resolution and execution, and model selection is per\-client and per\-tier behind a rate governor\.

A single*substrate*, parquet over DuckDB, is shared by the explorer, the report engine, and the live coder, so exploration imposes zero load on the client’s production database and the live path benefits from the same fast columnar access; an analytics backend abstraction prefers materialized parquet and falls back to read\-only generated SQL across the supported relational connectors\. Charts are rendered deterministically backend\-side from a typed specification rather than from free\-hand plotting code, which is what lets a report’s recipe reproduce identical visuals without an LLM\.

The engineering posture is uniform:*fail\-open at every boundary*and*additive behind flags*\. The skill framework, deep exploration, the deep report engine, and the kernel\-library hook are each gated, default to the prior behavior, and degrade to a no\-op on error \([Proposition˜1](https://arxiv.org/html/2608.28594#Thmproposition1)\); document shapes are backward\-compatible; and the system is covered by a large test suite plus a DuckDB\-oracle correctness harness\. The cost/latency trade\-off is made explicitly by path: the chat path uses a single small reference excerpt and is latency\-bounded, while exploration and report generation are quality\-first batch jobs that can afford full references and can wait out rate limits\.

## 9Case Study: Single\-Tenant Deployment

We report illustrative observations from one live tenant onboarded with Tally accounting data materialized to parquet\. We emphasize that these are*single\-deployment*figures, presented for transparency; we ran no controlled study and report no benchmark numbers\.

Table 1:Single\-tenant deployment evidence \(illustrative, observed from one live account\)\. These counts evidence end\-to\-end execution and convergence on one schema; they are*not*precision/recall of join discovery or report correctness, for which no ground truth was collected\.Qualitatively, the loop ran end to end with no human in the path: the value\-based detector recognized accounting data from distinctive ledger\-group strings even though the table names were opaque, exploration converged on all ten tables and proved twenty\-two joins, seven reports were authored and admitted under the verification gate, and roughly thirty domain questions were generated \(phrased in business vocabulary and grounded in the report KPIs and validated\-knowledge brief\) before the tenant ran a single query\. The published GST figure illustrates the trust contract: it appears*because*its evidence query reproduced it within tolerance, and views whose numbers could not be reproduced would have been suppressed rather than shown\.

## 10Discussion: Novelty and Positioning

The defensibility of this architecture is structural, and it survives counter\-positioning along three axes \([Table˜2](https://arxiv.org/html/2608.28594#S10.T2)\)\.

Table 2:Positioning along three axes: who supplies the question, what “proactive” means, and what guarantees the number\.#### Who supplies the question\.

Text\-to\-SQL and question\-first NLQ assume the user has a question; log\-based recommenders supply one but need prior logs and an active session, so they cold\-start on a fresh tenant\. We supply the question from a*portable domain\-expert prior*that exists on day one with no logs and no per\-client engineering, and that is guaranteed answerable because it is generated from the same auto\-explored, validated knowledge the pipeline answers against\.

#### What “proactive” means\.

Statistical insight miners surface what is statistically salient, not what a chartered accountant or marketing leader would act on; the most proactive metric\-driven product, Tableau Pulse, is bound to an analyst\-curated metric layer\[[1](https://arxiv.org/html/2608.28594#bib.bib58)\]\. We are proactive over an un\-curated schema \(for a domain a skill covers\) with*no*pre\-curated metric layer: the expert supplies the agenda from the skill’s templates and the explorer expands it into the cuts the actual columns support\.

#### What guarantees the number\.

Persona/expert injection is known to introduce domain\-correlated bias\[[51](https://arxiv.org/html/2608.28594#bib.bib21)\], and intrinsic self\-correction is unreliable\[[25](https://arxiv.org/html/2608.28594#bib.bib49)\]\. Our countermeasure is to fuse the opinionated expert with*execution\-grounded verification*: every published metric carries an evidence query that an independent, deterministic critic re\-executes and tolerance\-matches, and every join is proven on the data\. The expert sets the agenda; it cannot fabricate the figure\.

#### The honest differentiator\.

It is none of the commoditized parts in isolation: not NL\-to\-SQL, not anomaly detection, not skill packaging\. It is the*combination*, deployed and fail\-open: a portable domain\-expert layer threaded as a cross\-cutting concern through router, scout, coder, chart, narrator, the explorer, and both report engines; schema\-derived \(not metric\-bound\) agentic analysis; and end\-to\-end execution verification, instantiated as a closed proactive loop with a fully\-automatic onboarding chain\. We do not claim to invent skill packs \(modular agent skills predate and parallel this work\[[4](https://arxiv.org/html/2608.28594#bib.bib41)\]\); the novelty is the per\-tenant, schema\-auto\-selected analytics*specialization*fused with execution\-grounded trust\.

#### The trade\-off we accept\.

A curated metric layer\[[12](https://arxiv.org/html/2608.28594#bib.bib39),[32](https://arxiv.org/html/2608.28594#bib.bib40)\]encodes*definitional*correctness \(blessed business definitions\) that our schema\-derived agenda forgoes\. We trade per\-organization curation cost for portability and recover*arithmetic*\(not definitional\) trust via evidence re\-execution; bridging the two \(ingesting an existing metric layer as an additional skill input\) is natural future work\.

## 11Limitations and Threats to Validity

- •Single deployment, no controlled study\.Evidence is from one tenant; we report no user study and no public\-benchmark numbers\. The deployment figures are illustrative, and the two undeployed packs evidence the generality of the*structure*, not of results\.
- •Heuristic, human\-authored skills\.The catalogue is small and the packs are written by hand; auto\-detection is a heuristic that can miss or mis\-route on a truly generic schema, mitigated, but not eliminated, by explicit assignment and fail\-open\-to\-no\-skill \([Proposition˜1](https://arxiv.org/html/2608.28594#Thmproposition1)\)\.
- •Verification boundary\.[Equation˜7](https://arxiv.org/html/2608.28594#S3.E7)certifies that a published number matches its declared evidence query within tolerance; it does*not*certify that the metric’s business definition is semantically correct\. The guarantee is arithmetic faithfulness, not definitional correctness, and its failure mode is suppression, not a certificate of universal correctness\.
- •Persona bias\.An opinionated expert can bias framing\[[51](https://arxiv.org/html/2608.28594#bib.bib21)\]; deterministic grounding \(evidence re\-execution, typed formatting, deterministic rendering\) constrains the*numbers*but not necessarily the emphasis\.
- •Cost and staleness\.Exploration and report generation are token\-heavy offline jobs whose cost scales with tables and width; it is amortized by zero\-cost refresh, not free\. Compiled knowledge can go stale; it is regenerated on schema\-fingerprint change and via capped manual refresh, but is not continuously revalidated\.

## 12Conclusion

A pluggable domain\-expert skill abstraction and an offline knowledge\-compilation loop together flip enterprise data analysis from reactive question\-answering to a proactive, verified analyst: the system understands the data, authors the agenda, re\-verifies each published number against its evidence SQL, and tells the user what to ask, before the query box is ever used\. The architecture’s defensibility is not any one commoditized capability but their fusion: domain expertise as a cross\-cutting, auto\-selected, fail\-open concern, joined to execution\-grounded trust and closed into a proactive loop with an automatic onboarding chain\. Because expertise lives in self\-contained, deterministically selected, fail\-open packs, the natural trajectory is a*marketplace of domain experts*\(each the encoded playbook of a PhD\-level or veteran industry specialist\) that a tenant draws on automatically the moment data is connected\. Future work is that larger and partly auto\-authored catalogue,*semantic*\(not only numeric\) metric verification, continuous knowledge revalidation, and the rigorous user study this single\-deployment paper deliberately does not claim\.

## References

- \[1\]T\. \(Salesforce\)\(2025\)About tableau pulse\.Note:Tableau Help / Product Documentation[https://help\.tableau\.com/current/online/en\-us/pulse\_intro\.htm](https://help.tableau.com/current/online/en-us/pulse_intro.htm)Cited by:[§1](https://arxiv.org/html/2608.28594#S1.p2.1),[§10](https://arxiv.org/html/2608.28594#S10.SS0.SSS0.Px2.p1.1),[Table 2](https://arxiv.org/html/2608.28594#S10.T2.4.5.4.1),[§2](https://arxiv.org/html/2608.28594#S2.SS0.SSS0.Px8.p1.1)\.
- \[2\]T\. \(Salesforce\)\(2025\)The insights platform and insight types in tableau pulse\.Note:Tableau Help / Product Documentation[https://help\.tableau\.com/current/online/en\-us/pulse\_insights\_platform\_insight\_types\.htm](https://help.tableau.com/current/online/en-us/pulse_insights_platform_insight_types.htm)Cited by:[§1](https://arxiv.org/html/2608.28594#S1.p2.1),[Table 2](https://arxiv.org/html/2608.28594#S10.T2.4.5.4.1),[§2](https://arxiv.org/html/2608.28594#S2.SS0.SSS0.Px8.p1.1)\.
- \[3\]Z\. Abedjan, L\. Golab, and F\. Naumann\(2015\)Profiling relational data: a survey\.The VLDB Journal 24\(4\)\.Note:[https://link\.springer\.com/article/10\.1007/s00778\-015\-0389\-y](https://link.springer.com/article/10.1007/s00778-015-0389-y)Cited by:[§2](https://arxiv.org/html/2608.28594#S2.SS0.SSS0.Px7.p1.1)\.
- \[4\]Anthropic\(2025\)Equipping agents for the real world with agent skills\.Note:Anthropic engineering blog[https://www\.anthropic\.com/engineering/equipping\-agents\-for\-the\-real\-world\-with\-agent\-skills](https://www.anthropic.com/engineering/equipping-agents-for-the-real-world-with-agent-skills)Cited by:[§1](https://arxiv.org/html/2608.28594#S1.SS0.SSS0.Px2.p1.1),[§10](https://arxiv.org/html/2608.28594#S10.SS0.SSS0.Px4.p1.1),[Table 2](https://arxiv.org/html/2608.28594#S10.T2.4.6.5.1),[§2](https://arxiv.org/html/2608.28594#S2.SS0.SSS0.Px6.p1.1)\.
- \[5\]Anthropic\(2025\)Subagents \(create custom subagents\) — claude code docs\.Note:Claude Code documentation[https://code\.claude\.com/docs/en/sub\-agents](https://code.claude.com/docs/en/sub-agents)Cited by:[§1](https://arxiv.org/html/2608.28594#S1.SS0.SSS0.Px2.p1.1),[§2](https://arxiv.org/html/2608.28594#S2.SS0.SSS0.Px6.p1.1)\.
- \[6\]G\. Chatzopoulou, M\. Eirinaki, and N\. Polyzotis\(2009\)Query recommendations for interactive database exploration \(querie\)\.InSSDBM/DaWaK 2009 \(extended in VLDB 2011\),Note:[https://link\.springer\.com/chapter/10\.1007/978\-3\-642\-02279\-1\_2](https://link.springer.com/chapter/10.1007/978-3-642-02279-1_2)Cited by:[§1](https://arxiv.org/html/2608.28594#S1.p2.1),[Table 2](https://arxiv.org/html/2608.28594#S10.T2.4.3.2.1),[§2](https://arxiv.org/html/2608.28594#S2.SS0.SSS0.Px3.p1.1),[§7\.2](https://arxiv.org/html/2608.28594#S7.SS2.p1.1)\.
- \[7\]R\. Chaudhuri, R\. C, K\. DB, and A\. Agarwal\(2024\)Automated question generation on tabular data for conversational data exploration\.arXiv:2407\.12859\.Note:[https://arxiv\.org/abs/2407\.12859](https://arxiv.org/abs/2407.12859)Cited by:[§2](https://arxiv.org/html/2608.28594#S2.SS0.SSS0.Px3.p1.1)\.
- \[8\]X\. Chen, M\. Lin, N\. Scharli, and D\. Zhou\(2023\)Teaching large language models to self\-debug\.InICLR 2024 \(arXiv:2304\.05128\),Note:[https://arxiv\.org/abs/2304\.05128](https://arxiv.org/abs/2304.05128)Cited by:[§1](https://arxiv.org/html/2608.28594#S1.SS0.SSS0.Px2.p1.1),[§2](https://arxiv.org/html/2608.28594#S2.SS0.SSS0.Px7.p1.1),[§6\.1](https://arxiv.org/html/2608.28594#S6.SS1.p1.1)\.
- \[9\]G\. Cloud\(2025\)Conversational analytics in looker overview\.Note:Google Cloud Documentation[https://docs\.cloud\.google\.com/looker/docs/conversational\-analytics\-overview](https://docs.cloud.google.com/looker/docs/conversational-analytics-overview)Cited by:[§1](https://arxiv.org/html/2608.28594#S1.p1.1),[Table 2](https://arxiv.org/html/2608.28594#S10.T2.4.2.1.1),[§2](https://arxiv.org/html/2608.28594#S2.SS0.SSS0.Px8.p1.1)\.
- \[10\]K\. Cobbe, V\. Kosaraju, M\. Bavarian, M\. Chen, H\. Jun, L\. Kaiser, M\. Plappert, J\. Tworek, J\. Hilton, R\. Nakano, C\. Hesse, and J\. Schulman\(2021\)Training verifiers to solve math word problems\.arXiv:2110\.14168 \(OpenAI\)\.Note:[https://arxiv\.org/abs/2110\.14168](https://arxiv.org/abs/2110.14168)Cited by:[§2](https://arxiv.org/html/2608.28594#S2.SS0.SSS0.Px7.p1.1)\.
- \[11\]Z\. Cui, S\. K\. Badam, M\. A\. Yalçin, and N\. Elmqvist\(2019\)DataSite: proactive visual data exploration with computation of insight\-based recommendations\.Information Visualization \(SAGE\) 18\(2\)\.Note:[https://journals\.sagepub\.com/doi/10\.1177/1473871618806555](https://journals.sagepub.com/doi/10.1177/1473871618806555)Cited by:[§2](https://arxiv.org/html/2608.28594#S2.SS0.SSS0.Px4.p1.1)\.
- \[12\]dbt Labs\(2024\)Dbt semantic layer \(powered by metricflow\) — documentation\.Note:Vendor/engineering documentation[https://docs\.getdbt\.com/docs/use\-dbt\-semantic\-layer/dbt\-sl](https://docs.getdbt.com/docs/use-dbt-semantic-layer/dbt-sl)Cited by:[§10](https://arxiv.org/html/2608.28594#S10.SS0.SSS0.Px5.p1.1),[§2](https://arxiv.org/html/2608.28594#S2.SS0.SSS0.Px6.p1.1)\.
- \[13\]C\. Demiralp, P\. J\. Haas, S\. Parthasarathy, and T\. Pedapati\(2017\)Foresight: recommending visual insights\.InVLDB 2017 \(PVLDB 10:12\),Note:[https://dl\.acm\.org/doi/10\.14778/3137765\.3137813](https://dl.acm.org/doi/10.14778/3137765.3137813)Cited by:[§1](https://arxiv.org/html/2608.28594#S1.p2.1),[Table 2](https://arxiv.org/html/2608.28594#S10.T2.4.4.3.1),[§2](https://arxiv.org/html/2608.28594#S2.SS0.SSS0.Px4.p1.1)\.
- \[14\]V\. Dibia\(2023\)LIDA: a tool for automatic generation of grammar\-agnostic visualizations and infographics using large language models\.InACL 2023 \(System Demonstrations\),Note:[https://arxiv\.org/abs/2303\.02927](https://arxiv.org/abs/2303.02927)Cited by:[§2](https://arxiv.org/html/2608.28594#S2.SS0.SSS0.Px5.p1.1)\.
- \[15\]R\. Ding, S\. Han, Y\. Xu, H\. Zhang, and D\. Zhang\(2019\)QuickInsights: quick and automatic discovery of insights from multi\-dimensional data\.InSIGMOD 2019,Note:[https://dl\.acm\.org/doi/10\.1145/3299869\.3314037](https://dl.acm.org/doi/10.1145/3299869.3314037)Cited by:[§1](https://arxiv.org/html/2608.28594#S1.p2.1),[Table 2](https://arxiv.org/html/2608.28594#S10.T2.4.4.3.1),[§2](https://arxiv.org/html/2608.28594#S2.SS0.SSS0.Px4.p1.1),[§7\.2](https://arxiv.org/html/2608.28594#S7.SS2.p1.1)\.
- \[16\]T\. Gao, M\. Dontcheva, E\. Adar, Z\. Liu, and K\. Karahalios\(2015\)DataTone: managing ambiguity in natural language interfaces for data visualization\.InUIST 2015,Note:[https://dl\.acm\.org/doi/10\.1145/2807442\.2807478](https://dl.acm.org/doi/10.1145/2807442.2807478)Cited by:[§2](https://arxiv.org/html/2608.28594#S2.SS0.SSS0.Px2.p1.1)\.
- \[17\]Y\. Gao, Y\. Xiong, X\. Gao, K\. Jia, J\. Pan, Y\. Bi, Y\. Dai, J\. Sun, M\. Wang, and H\. Wang\(2024\)Retrieval\-augmented generation for large language models: a survey\.arXiv:2312\.10997\.Note:[https://arxiv\.org/abs/2312\.10997](https://arxiv.org/abs/2312.10997)Cited by:[§2](https://arxiv.org/html/2608.28594#S2.SS0.SSS0.Px6.p1.1)\.
- \[18\]Z\. Gou, Z\. Shao, Y\. Gong, Y\. Shen, Y\. Yang, N\. Duan, and W\. Chen\(2023\)CRITIC: large language models can self\-correct with tool\-interactive critiquing\.InICLR 2024 \(arXiv:2305\.11738\),Note:[https://arxiv\.org/abs/2305\.11738](https://arxiv.org/abs/2305.11738)Cited by:[§1](https://arxiv.org/html/2608.28594#S1.SS0.SSS0.Px2.p1.1),[§2](https://arxiv.org/html/2608.28594#S2.SS0.SSS0.Px7.p1.1),[§6\.1](https://arxiv.org/html/2608.28594#S6.SS1.p1.1)\.
- \[19\]S\. Hong, Y\. Lin, B\. Liu, B\. Liu, B\. Wu, D\. Li, J\. Chen, J\. Zhang, and J\. Wang\(2024\)Data interpreter: an llm agent for data science\.InarXiv:2402\.18679 \(later ACL 2025 Findings\),Note:[https://arxiv\.org/abs/2402\.18679](https://arxiv.org/abs/2402.18679)Cited by:[§2](https://arxiv.org/html/2608.28594#S2.SS0.SSS0.Px5.p1.1)\.
- \[20\]S\. Hong, M\. Zhuge, J\. Chen, X\. Zheng, Y\. Cheng, C\. Zhang, J\. Wang, Z\. Wang, S\. K\. S\. Yau, Z\. Lin, L\. Zhou, C\. Ran, L\. Xiao, C\. Wu, and J\. Schmidhuber\(2024\)MetaGPT: meta programming for a multi\-agent collaborative framework\.InICLR 2024,Note:[https://arxiv\.org/abs/2308\.00352](https://arxiv.org/abs/2308.00352)Cited by:[§2](https://arxiv.org/html/2608.28594#S2.SS0.SSS0.Px5.p1.1)\.
- \[21\]Z\. Hong, Z\. Yuan, H\. Chen, Q\. Zhang, F\. Huang, and X\. Huang\(2024\)Knowledge\-to\-sql: enhancing sql generation with data expert llm\.InFindings of ACL 2024,Note:[https://aclanthology\.org/2024\.findings\-acl\.653/](https://aclanthology.org/2024.findings-acl.653/)Cited by:[§2](https://arxiv.org/html/2608.28594#S2.SS0.SSS0.Px6.p1.1)\.
- \[22\]Z\. Hong, Z\. Yuan, Q\. Zhang, H\. Chen, J\. Dong, F\. Huang, and X\. Huang\(2024\)Next\-generation database interfaces: a survey of llm\-based text\-to\-sql\.arXiv:2406\.08426 \(later IEEE TKDE 2025\)\.Note:[https://arxiv\.org/abs/2406\.08426](https://arxiv.org/abs/2406.08426)Cited by:[§1](https://arxiv.org/html/2608.28594#S1.p1.1),[§2](https://arxiv.org/html/2608.28594#S2.SS0.SSS0.Px1.p1.1)\.
- \[23\]E\. Horvitz\(1999\)Principles of mixed\-initiative user interfaces\.InCHI 1999,Note:[https://dl\.acm\.org/doi/10\.1145/302979\.303030](https://dl.acm.org/doi/10.1145/302979.303030)Cited by:[§2](https://arxiv.org/html/2608.28594#S2.SS0.SSS0.Px4.p1.1)\.
- \[24\]X\. Hu, Z\. Zhao, S\. Wei, Z\. Chai, Q\. Ma, G\. Wang, X\. Wang, J\. Su, J\. Xu, M\. Zhu, Y\. Hu, Z\. Wei, F\. Wu, Y\. Yang, and K\. Kuang\(2024\)InfiAgent\-dabench: evaluating agents on data analysis tasks\.InICML 2024,Note:[https://arxiv\.org/abs/2401\.05507](https://arxiv.org/abs/2401.05507)Cited by:[§2](https://arxiv.org/html/2608.28594#S2.SS0.SSS0.Px5.p1.1)\.
- \[25\]J\. Huang, X\. Chen, S\. Mishra, H\. S\. Zheng, A\. W\. Yu, X\. Song, and D\. Zhou\(2023\)Large language models cannot self\-correct reasoning yet\.InICLR 2024 \(arXiv:2310\.01798\),Note:[https://arxiv\.org/abs/2310\.01798](https://arxiv.org/abs/2310.01798)Cited by:[§10](https://arxiv.org/html/2608.28594#S10.SS0.SSS0.Px3.p1.1),[§2](https://arxiv.org/html/2608.28594#S2.SS0.SSS0.Px7.p1.1),[§5\.1](https://arxiv.org/html/2608.28594#S5.SS1.SSS0.Px1.p1.1)\.
- \[26\]F\. Lei, J\. Chen, Y\. Ye, R\. Cao, D\. Shin, H\. Su, Z\. Suo, H\. Gao, W\. Hu, P\. Yin, V\. Zhong, C\. Xiong, R\. Sun, Q\. Liu, S\. Wang, and T\. Yu\(2024\)Spider 2\.0: evaluating language models on real\-world enterprise text\-to\-sql workflows\.InICLR 2025 \(arXiv:2411\.07763\),Note:[https://arxiv\.org/abs/2411\.07763](https://arxiv.org/abs/2411.07763)Cited by:[§2](https://arxiv.org/html/2608.28594#S2.SS0.SSS0.Px1.p1.1)\.
- \[27\]C\. Lennerholt, J\. van Laere, and E\. Söderström\(2021\)User\-related challenges of self\-service business intelligence\.Information Systems Management, 38\(4\)\.Note:[https://www\.tandfonline\.com/doi/full/10\.1080/10580530\.2020\.1814458](https://www.tandfonline.com/doi/full/10.1080/10580530.2020.1814458)Cited by:[§1](https://arxiv.org/html/2608.28594#S1.p1.1)\.
- \[28\]P\. Lewis, E\. Perez, A\. Piktus, F\. Petroni, V\. Karpukhin, N\. Goyal, H\. Küttler, M\. Lewis, W\. Yih, T\. Rocktäschel, S\. Riedel, and D\. Kiela\(2020\)Retrieval\-augmented generation for knowledge\-intensive nlp tasks\.InNeurIPS 2020,Note:[https://proceedings\.neurips\.cc/paper/2020/hash/6b493230205f780e1bc26945df7481e5\-Abstract\.html](https://proceedings.neurips.cc/paper/2020/hash/6b493230205f780e1bc26945df7481e5-Abstract.html)Cited by:[§2](https://arxiv.org/html/2608.28594#S2.SS0.SSS0.Px6.p1.1)\.
- \[29\]F\. Li and H\. V\. Jagadish\(2014\)NaLIR: an interactive natural language interface for querying relational databases\.InSIGMOD 2014,Note:[http://dbgroup\.eecs\.umich\.edu/files/SIGMOD14LFb\.pdf](http://dbgroup.eecs.umich.edu/files/SIGMOD14LFb.pdf)Cited by:[§2](https://arxiv.org/html/2608.28594#S2.SS0.SSS0.Px2.p1.1)\.
- \[30\]J\. Li, B\. Hui, G\. Qu, J\. Yang, B\. Li, B\. Li, B\. Wang, B\. Qin, R\. Geng, N\. Huo, X\. Zhou, C\. Ma, G\. Li, K\. C\. C\. Chang, F\. Huang, R\. Cheng, and Y\. Li\(2023\)Can llm already serve as a database interface? a big bench for large\-scale database grounded text\-to\-sqls \(bird\)\.InNeurIPS 2023 Datasets & Benchmarks,Note:[https://arxiv\.org/abs/2305\.03111](https://arxiv.org/abs/2305.03111)Cited by:[§1](https://arxiv.org/html/2608.28594#S1.p1.1),[Table 2](https://arxiv.org/html/2608.28594#S10.T2.4.2.1.1),[§2](https://arxiv.org/html/2608.28594#S2.SS0.SSS0.Px1.p1.1)\.
- \[31\]H\. Lightman, V\. Kosaraju, Y\. Burda, H\. Edwards, B\. Baker, T\. Lee, J\. Leike, J\. Schulman, I\. Sutskever, and K\. Cobbe\(2023\)Let’s verify step by step\.InICLR 2024 \(arXiv:2305\.20050\),Note:[https://arxiv\.org/abs/2305\.20050](https://arxiv.org/abs/2305.20050)Cited by:[§2](https://arxiv.org/html/2608.28594#S2.SS0.SSS0.Px7.p1.1)\.
- \[32\]G\. /\. Looker\(2025\)Looker modeling \(lookml\) — google cloud\.Note:Vendor/engineering documentation[https://cloud\.google\.com/looker\-modeling](https://cloud.google.com/looker-modeling)Cited by:[§10](https://arxiv.org/html/2608.28594#S10.SS0.SSS0.Px5.p1.1),[§2](https://arxiv.org/html/2608.28594#S2.SS0.SSS0.Px6.p1.1)\.
- \[33\]Y\. Lu, S\. Yang, C\. Qian, G\. Chen, Q\. Luo, Y\. Wu, H\. Wang, X\. Cong, Z\. Zhang, Y\. Lin, W\. Liu, Y\. Wang, Z\. Liu, F\. Liu, and M\. Sun\(2024\)Proactive agent: shifting llm agents from reactive responses to active assistance\.arXiv:2410\.12361\.Note:[https://arxiv\.org/abs/2410\.12361](https://arxiv.org/abs/2410.12361)Cited by:[§2](https://arxiv.org/html/2608.28594#S2.SS0.SSS0.Px4.p1.1)\.
- \[34\]P\. Ma, R\. Ding, S\. Han, and D\. Zhang\(2021\)MetaInsight: automatic discovery of structured knowledge for exploratory data analysis\.InSIGMOD 2021,Note:[https://dl\.acm\.org/doi/10\.1145/3448016\.3457267](https://dl.acm.org/doi/10.1145/3448016.3457267)Cited by:[§2](https://arxiv.org/html/2608.28594#S2.SS0.SSS0.Px4.p1.1)\.
- \[35\]P\. Ma, R\. Ding, S\. Wang, S\. Han, and D\. Zhang\(2023\)InsightPilot: an llm\-empowered automated data exploration system\.InACL 2023 \(System Demonstrations\),Note:[https://arxiv\.org/abs/2304\.00477](https://arxiv.org/abs/2304.00477)Cited by:[§2](https://arxiv.org/html/2608.28594#S2.SS0.SSS0.Px5.p1.1)\.
- \[36\]A\. Madaan, N\. Tandon, P\. Gupta, S\. Hallinan, L\. Gao, S\. Wiegreffe, U\. Alon, N\. Dziri, S\. Prabhumoye, Y\. Yang, S\. Gupta, B\. P\. Majumder, K\. Hermann, S\. Welleck, A\. Yazdanbakhsh, and P\. Clark\(2023\)Self\-refine: iterative refinement with self\-feedback\.InNeurIPS 2023 \(arXiv:2303\.17651\),Note:[https://arxiv\.org/abs/2303\.17651](https://arxiv.org/abs/2303.17651)Cited by:[§2](https://arxiv.org/html/2608.28594#S2.SS0.SSS0.Px7.p1.1)\.
- \[37\]Microsoft\(2025\)Generate data insights on your semantic model automatically \(quick insights\) — power bi\.Note:Microsoft Learn / Power BI Documentation[https://learn\.microsoft\.com/en\-us/power\-bi/create\-reports/service\-insights](https://learn.microsoft.com/en-us/power-bi/create-reports/service-insights)Cited by:[§1](https://arxiv.org/html/2608.28594#S1.p2.1),[Table 2](https://arxiv.org/html/2608.28594#S10.T2.4.4.3.1),[§2](https://arxiv.org/html/2608.28594#S2.SS0.SSS0.Px8.p1.1)\.
- \[38\]S\. Monadjemi, Y\. Guo, K\. Xu, A\. Endert, and A\. Crisan\(2025\)A scoping review of mixed\-initiative visual analytics in the automation renaissance\.arXiv:2509\.19152\.Note:[https://arxiv\.org/abs/2509\.19152](https://arxiv.org/abs/2509.19152)Cited by:[§2](https://arxiv.org/html/2608.28594#S2.SS0.SSS0.Px4.p1.1)\.
- \[39\]C\. Packer, S\. Wooders, K\. Lin, V\. Fang, S\. G\. Patil, I\. Stoica, and J\. E\. Gonzalez\(2023\)MemGPT: towards llms as operating systems\.arXiv:2310\.08560\.Note:[https://arxiv\.org/abs/2310\.08560](https://arxiv.org/abs/2310.08560)Cited by:[§2](https://arxiv.org/html/2608.28594#S2.SS0.SSS0.Px7.p1.1)\.
- \[40\]L\. Pan, M\. Saxon, W\. Xu, D\. Nathani, X\. Wang, and W\. Y\. Wang\(2024\)Automatically correcting large language models: surveying the landscape of diverse automated correction strategies\.TACL, vol\. 12\.Note:[https://aclanthology\.org/2024\.tacl\-1\.27/](https://aclanthology.org/2024.tacl-1.27/)Cited by:[§2](https://arxiv.org/html/2608.28594#S2.SS0.SSS0.Px7.p1.1)\.
- \[41\]J\. S\. Park, J\. C\. O’Brien, C\. J\. Cai, M\. R\. Morris, P\. Liang, and M\. S\. Bernstein\(2023\)Generative agents: interactive simulacra of human behavior\.InUIST 2023 \(arXiv:2304\.03442\),Note:[https://arxiv\.org/abs/2304\.03442](https://arxiv.org/abs/2304.03442)Cited by:[§2](https://arxiv.org/html/2608.28594#S2.SS0.SSS0.Px7.p1.1)\.
- \[42\]T\. Schick, J\. Dwivedi\-Yu, R\. Dessì, R\. Raileanu, M\. Lomeli, L\. Zettlemoyer, N\. Cancedda, and T\. Scialom\(2023\)Toolformer: language models can teach themselves to use tools\.InNeurIPS 2023,Note:[https://arxiv\.org/abs/2302\.04761](https://arxiv.org/abs/2302.04761)Cited by:[§2](https://arxiv.org/html/2608.28594#S2.SS0.SSS0.Px6.p1.1)\.
- \[43\]A\. W\. Services\(2021\)New — amazon quicksight q answers natural\-language questions about business data\.Note:AWS News Blog[https://aws\.amazon\.com/blogs/aws/amazon\-quicksight\-q\-to\-answer\-ad\-hoc\-business\-questions/](https://aws.amazon.com/blogs/aws/amazon-quicksight-q-to-answer-ad-hoc-business-questions/)Cited by:[§1](https://arxiv.org/html/2608.28594#S1.p1.1),[Table 2](https://arxiv.org/html/2608.28594#S10.T2.4.2.1.1),[§2](https://arxiv.org/html/2608.28594#S2.SS0.SSS0.Px8.p1.1)\.
- \[44\]A\. W\. Services\(2025\)Amazon quicksight launches amazon q in embedded quicksight\.Note:AWS What’s New[https://aws\.amazon\.com/about\-aws/whats\-new/2025/04/amazon\-quicksight\-q\-embedded/](https://aws.amazon.com/about-aws/whats-new/2025/04/amazon-quicksight-q-embedded/)Cited by:[§2](https://arxiv.org/html/2608.28594#S2.SS0.SSS0.Px8.p1.1)\.
- \[45\]V\. Setlur, S\. E\. Battersby, M\. Tory, R\. Gossweiler, and A\. X\. Chang\(2016\)Eviza: a natural language interface for visual analysis\.InUIST 2016,Note:[https://www\.researchgate\.net/publication/309222635\_Eviza\_A\_Natural\_Language\_Interface\_for\_Visual\_Analysis](https://www.researchgate.net/publication/309222635_Eviza_A_Natural_Language_Interface_for_Visual_Analysis)Cited by:[§2](https://arxiv.org/html/2608.28594#S2.SS0.SSS0.Px2.p1.1)\.
- \[46\]L\. Shi, Z\. Tang, N\. Zhang, X\. Zhang, and Z\. Yang\(2024\)A survey on employing large language models for text\-to\-sql tasks\.arXiv:2407\.15186\.Note:[https://arxiv\.org/abs/2407\.15186](https://arxiv.org/abs/2407.15186)Cited by:[§2](https://arxiv.org/html/2608.28594#S2.SS0.SSS0.Px1.p1.1)\.
- \[47\]N\. Shinn, F\. Cassano, E\. Berman, A\. Gopinath, K\. Narasimhan, and S\. Yao\(2023\)Reflexion: language agents with verbal reinforcement learning\.InNeurIPS 2023,Note:[https://arxiv\.org/abs/2303\.11366](https://arxiv.org/abs/2303.11366)Cited by:[§1](https://arxiv.org/html/2608.28594#S1.SS0.SSS0.Px2.p1.1),[§2](https://arxiv.org/html/2608.28594#S2.SS0.SSS0.Px5.p1.1),[§5](https://arxiv.org/html/2608.28594#S5.p1.1)\.
- \[48\]T\. Siddiqui, A\. Kim, J\. Lee, K\. Karahalios, and A\. Parameswaran\(2017\)Effortless data exploration with zenvisage: an expressive and interactive visual analytics system\.InVLDB 2017 \(PVLDB 10:4\),Note:[https://dl\.acm\.org/doi/10\.14778/3025111\.3025126](https://dl.acm.org/doi/10.14778/3025111.3025126)Cited by:[§2](https://arxiv.org/html/2608.28594#S2.SS0.SSS0.Px4.p1.1)\.
- \[49\]ThoughtSpot\(2025\)ThoughtSpot augmented analytics: spotiq\.Note:ThoughtSpot Product Page[https://www\.thoughtspot\.com/product/analytics/spotiq](https://www.thoughtspot.com/product/analytics/spotiq)Cited by:[§1](https://arxiv.org/html/2608.28594#S1.p2.1),[Table 2](https://arxiv.org/html/2608.28594#S10.T2.4.4.3.1),[§2](https://arxiv.org/html/2608.28594#S2.SS0.SSS0.Px8.p1.1)\.
- \[50\]D\. Trends and A\. \(DBTA\)\(2024\)Salesforce debuts tableau einstein to autonomously discover and surface insights\.Note:DBTA News[https://www\.dbta\.com/Editorial/News\-Flashes/Salesforce\-Debuts\-Tableau\-Einstein\-to\-Autonomously\-Discover\-and\-Surface\-Insights\-165944\.aspx](https://www.dbta.com/Editorial/News-Flashes/Salesforce-Debuts-Tableau-Einstein-to-Autonomously-Discover-and-Surface-Insights-165944.aspx)Cited by:[§2](https://arxiv.org/html/2608.28594#S2.SS0.SSS0.Px8.p1.1)\.
- \[51\]Y\. Tseng, Y\. Huang, T\. Hsiao, W\. Chen, C\. Huang, Y\. Meng, and Y\. Chen\(2024\)Two tales of persona in llms: a survey of role\-playing and personalization\.InEMNLP 2024 Findings,Note:[https://arxiv\.org/abs/2406\.01171](https://arxiv.org/abs/2406.01171)Cited by:[§10](https://arxiv.org/html/2608.28594#S10.SS0.SSS0.Px3.p1.1),[4th item](https://arxiv.org/html/2608.28594#S11.I1.i4.p1.1),[§2](https://arxiv.org/html/2608.28594#S2.SS0.SSS0.Px7.p1.1)\.
- \[52\]M\. Vartak, S\. Rahman, S\. Madden, A\. Parameswaran, and N\. Polyzotis\(2015\)SeeDB: efficient data\-driven visualization recommendations to support visual analytics\.InVLDB 2015 \(PVLDB 8:13\),Note:[https://dl\.acm\.org/doi/10\.14778/2831360\.2831371](https://dl.acm.org/doi/10.14778/2831360.2831371)Cited by:[§1](https://arxiv.org/html/2608.28594#S1.p2.1),[Table 2](https://arxiv.org/html/2608.28594#S10.T2.4.4.3.1),[§2](https://arxiv.org/html/2608.28594#S2.SS0.SSS0.Px4.p1.1)\.
- \[53\]G\. Wang, Y\. Xie, Y\. Jiang, A\. Mandlekar, C\. Xiao, Y\. Zhu, L\. Fan, and A\. Anandkumar\(2023\)Voyager: an open\-ended embodied agent with large language models\.InTMLR 2024 \(arXiv:2305\.16291\),Note:[https://arxiv\.org/abs/2305\.16291](https://arxiv.org/abs/2305.16291)Cited by:[Table 2](https://arxiv.org/html/2608.28594#S10.T2.4.6.5.1),[§2](https://arxiv.org/html/2608.28594#S2.SS0.SSS0.Px7.p1.1)\.
- \[54\]X\. Wang, F\. Cheng, Y\. Wang, K\. Xu, J\. Long, H\. Lu, and H\. Qu\(2022\)Interactive data analysis with next\-step natural language query recommendation\.arXiv:2201\.04868\.Note:[https://arxiv\.org/abs/2201\.04868](https://arxiv.org/abs/2201.04868)Cited by:[§1](https://arxiv.org/html/2608.28594#S1.p2.1),[Table 2](https://arxiv.org/html/2608.28594#S10.T2.4.3.2.1),[§2](https://arxiv.org/html/2608.28594#S2.SS0.SSS0.Px3.p1.1),[§7\.2](https://arxiv.org/html/2608.28594#S7.SS2.p1.1)\.
- \[55\]X\. Wang, J\. Wei, D\. Schuurmans, Q\. Le, E\. Chi, S\. Narang, A\. Chowdhery, and D\. Zhou\(2022\)Self\-consistency improves chain of thought reasoning in language models\.InICLR 2023 \(arXiv:2203\.11171\),Note:[https://arxiv\.org/abs/2203\.11171](https://arxiv.org/abs/2203.11171)Cited by:[§2](https://arxiv.org/html/2608.28594#S2.SS0.SSS0.Px7.p1.1)\.
- \[56\]K\. Wongsuphasawat, D\. Moritz, A\. Anand, J\. Mackinlay, B\. Howe, and J\. Heer\(2016\)Voyager: exploratory analysis via faceted browsing of visualization recommendations\.IEEE TVCG 22\(1\)\.Note:[https://dl\.acm\.org/doi/10\.1109/TVCG\.2015\.2467191](https://dl.acm.org/doi/10.1109/TVCG.2015.2467191)Cited by:[§2](https://arxiv.org/html/2608.28594#S2.SS0.SSS0.Px4.p1.1)\.
- \[57\]K\. Wongsuphasawat, Z\. Qu, D\. Moritz, R\. Chang, F\. Ouk, A\. Anand, J\. Mackinlay, B\. Howe, and J\. Heer\(2017\)Voyager 2: augmenting visual analysis with partial view specifications\.InCHI 2017,Note:[https://dl\.acm\.org/doi/10\.1145/3025453\.3025768](https://dl.acm.org/doi/10.1145/3025453.3025768)Cited by:[§2](https://arxiv.org/html/2608.28594#S2.SS0.SSS0.Px4.p1.1)\.
- \[58\]Q\. Wu, G\. Bansal, J\. Zhang, Y\. Wu, B\. Li, E\. Zhu, L\. Jiang, X\. Zhang, S\. Zhang, J\. Liu, A\. Awadallah, R\. W\. White, D\. Burger, and C\. Wang\(2023\)AutoGen: enabling next\-gen llm applications via multi\-agent conversation framework\.InarXiv:2308\.08155 \(later COLM\),Note:[https://arxiv\.org/abs/2308\.08155](https://arxiv.org/abs/2308.08155)Cited by:[§2](https://arxiv.org/html/2608.28594#S2.SS0.SSS0.Px5.p1.1)\.
- \[59\]S\. Yao, J\. Zhao, D\. Yu, N\. Du, I\. Shafran, K\. Narasimhan, and Y\. Cao\(2023\)ReAct: synergizing reasoning and acting in language models\.InICLR 2023,Note:[https://arxiv\.org/abs/2210\.03629](https://arxiv.org/abs/2210.03629)Cited by:[§1](https://arxiv.org/html/2608.28594#S1.SS0.SSS0.Px2.p1.1),[§2](https://arxiv.org/html/2608.28594#S2.SS0.SSS0.Px5.p1.1),[§5](https://arxiv.org/html/2608.28594#S5.p1.1)\.
- \[60\]T\. Yu, R\. Zhang, K\. Yang, M\. Yasunaga, D\. Wang, Z\. Li, J\. Ma, I\. Li, Q\. Yao, S\. Roman, Z\. Zhang, and D\. Radev\(2018\)Spider: a large\-scale human\-labeled dataset for complex and cross\-domain semantic parsing and text\-to\-sql task\.InEMNLP 2018,Note:[https://aclanthology\.org/D18\-1425/](https://aclanthology.org/D18-1425/)Cited by:[§1](https://arxiv.org/html/2608.28594#S1.p1.1),[Table 2](https://arxiv.org/html/2608.28594#S10.T2.4.2.1.1),[§2](https://arxiv.org/html/2608.28594#S2.SS0.SSS0.Px1.p1.1)\.
- \[61\]R\. Zhang and M\. Elhamod\(2025\)Data\-to\-dashboard: multi\-agent llm framework for insightful visualization in enterprise analytics\.arXiv:2505\.23695\.Note:[https://arxiv\.org/abs/2505\.23695](https://arxiv.org/abs/2505.23695)Cited by:[§2](https://arxiv.org/html/2608.28594#S2.SS0.SSS0.Px5.p1.1)\.
- \[62\]Y\. Zhao, X\. Shu, L\. Fan, L\. Gao, Y\. Zhang, and S\. Chen\(2025\)ProactiveVA: proactive visual analytics with llm\-based ui agent\.arXiv:2507\.18165\.Note:[https://arxiv\.org/abs/2507\.18165](https://arxiv.org/abs/2507.18165)Cited by:[§2](https://arxiv.org/html/2608.28594#S2.SS0.SSS0.Px4.p1.1)\.
- \[63\]L\. Zheng, W\. Chiang, Y\. Sheng, S\. Zhuang, Z\. Wu, Y\. Zhuang, Z\. Lin, Z\. Li, D\. Li, E\. P\. Xing, H\. Zhang, J\. E\. Gonzalez, and I\. Stoica\(2023\)Judging llm\-as\-a\-judge with mt\-bench and chatbot arena\.InNeurIPS 2023 Datasets & Benchmarks \(arXiv:2306\.05685\),Note:[https://arxiv\.org/abs/2306\.05685](https://arxiv.org/abs/2306.05685)Cited by:[§2](https://arxiv.org/html/2608.28594#S2.SS0.SSS0.Px7.p1.1)\.