We’re excited to introduce the Retrieval Harness in LlamaParse - which is the 2026 version of RAG over documents

Generalized agents need the right set of tools to scalably search and read through an arbitrary corpus of data (from 10 docs to 1m+ docs). They can already demonstrate great retrieval performance over a local filesystem, need a proper backend for a large collection of managed data. The Retrieval Harness exposes a diverse set of tools for various needs:

1. Hybrid Retrieval: Combine vector search with keyword search, let the agent set the alpha value to toggle between the two
2. List Files: a scalable version of ls to list files within an index
3. File Grep: enable regex search within a given file
4. File Read: Allow agents to read a subsection from an existing document.

The agent can choose to interleave any sequence of these tools in order to complete a variety of tasks, from simple to hard.

Come check it out!

Blog: https://llamaindex.ai/blog/announcing-retrieval-harness?utm_medium=socials&utm_source=twitter&utm_campaign=2026-jun-…

Sign up to LlamaParse: https://cloud.llamaindex.ai

LlamaParse Retrieval Harness: Filesystem Primitives for AI Agents

Source: https://www.llamaindex.ai/blog/announcing-retrieval-harness?utm_medium=socials&utm_source=twitter&utm_campaign=2026-jun- When we launched LlamaIndex as an open-source project, our focus was on standardizing the core primitives of RAG: chunking, embedding, indexing, and retrieval. For basic question-answering workloads, that foundational blueprint worked perfectly.

But enterprise agents have completely outgrown it.

Traditional RAG treats data access as a static, one-shot preprocessing step. It pulls a handful of context fragments, bundles them into a prompt window, and blindly hopes for the best. Autonomous agents cannot navigate an unstructured corpus through a fuzzy semantic search bar. They require deterministic, systems-level utilities to actively interrogate, verify, and traverse documents in real time.

Today we are expanding LlamaParse Index to include: aRetrieval Harnessthat provides filesystem primitives for document traversal,visual layout preservation,managed indexes, andpipeline observability.

Pure semantic search dead-ends the moment an answer spans across arbitrary chunk boundaries. When that happens, trying to recover by letting an agent brute-force crawl a directory file by file completely torches your token budgets and latency constraints.

The Retrieval Harness solves this by exposing the underlying corpus as a set of filesystem-style tools that your agents can natively call:

• **Hybrid Retrieve:**A high-recall first pass that combines vector similarity with keyword search and out-of-the-box reranking to instantly narrow down the agent’s initial search space.
• **List Files:**File discovery. Allows agents to explicitly list the files contained within an index, giving them a clear map of the available document structure.
• **File Grep:**Server-side regex scanning on a targeted file. If an agent needs to isolate a specific serial number, error code, or exact phrase within a file, it does not waste tokens loading irrelevant semantic chunks. It executes a regex query directly against that file’s parsed text.
• **File Read:**Overcoming chunk fragmentation. When a top-k chunk cuts context off mid-sentence, the agent invokes a direct read API to pull the surrounding file context and seamlessly recover the missing data.

Visual Layout Preservation

For documents where text extraction alone isn’t enough, we now capture page screenshots at parse time and link them directly to their source chunks.

Financial tables, regulatory forms, and architectural diagrams where layout carries the structural meaning lose critical context when flattened into raw text strings. When the retrieved text isn’t sufficient to resolve an ambiguity, the agent can pull the actual rendered page exactly as LlamaParse processed it. This keeps the agent’s reasoning grounded directly in the source visual layout, preventing hallucinations on dense tables or multi-column documents.

Managed infrastructure

Setting up a production indexing pipeline means making a lot of decisions before you write a single line of application code. Data sources, embedding models, vector stores, sync logic, retrieval configuration. Then running it means dealing with rate limits, API failures, and pipeline breaks that are hard to diagnose.

LlamaParse Index now orchestrates this infrastructure layer natively out of the box. You simply connect your documents, and the platform provisions an optimized, production-grade baseline automatically—wiping out the manual setup drag so you can focus entirely on your application.

• **Incremental Sync:**We track which files have changed and only processes those. A folder with 1000 documents that receives 50 new files runs 50 files through the pipeline on the next sync. Parse costs and latency scale with actual document activity, not folder size.
• **Data Portability:**If you need to bring your own vector store or embedding model, the parsed outputs are downloadable.

Pipeline Observability

Production retrieval pipelines fail in non-obvious ways. A sync completes but the chunks that should have made it into the index didn’t. A stage reports green while the one after it has silently stalled. By the time retrieval quality degrades, the failure is several steps removed from where it actually happened.

We have built native, stage-by-stage pipeline tracking straight intoLlamaParse Index. Each stage of the pipeline has its own status and file count. When a sync completes but files are missing from the index, the stage counts show where they stopped. You know whether it’s an ingest failure or a workflow failure. You fix it instead of reconstructing what happened.

Integration and Availability

The Retrieval Harness and other updates are now available in beta across all paid tiers. All filesystem tools are exposed as lightweight API schemas that can be wired directly into your existing LLM orchestration frameworks or tool-calling loops.

👉Dive into the documentationto view the API specs, or initialize your first managed index directly from the LlamaIndex dashboard.

LlamaIndex 🦙 (@llama_index): Semantic search alone doesn’t cut it. Neither does brute-force grep. Agents need both.

Today we’re shipping the Retrieval Harness in LlamaParse Index: semantic search, server-side grep, and file-level navigation working together in a single agent reasoning loop. 🦙🌤️

Grep a