Mistral Agentic Search Release

Mistral AI has announced Agentic Search, a multi-step retrieval system designed to help enterprises navigate complex, dense, and sensitive data sources. Unlike traditional retrieval-augmented generation (RAG), Agentic Search allows models to iteratively search, inspect, and verify information across multiple documents, significantly increasing correctness in data-heavy domains.

Overcoming the Limitations of One-Shot RAG

Traditional one-shot RAG retrieves a fixed set of text chunks and attempts to answer in a single pass. Mistral identifies three primary failure points in this approach when dealing with complex data:

Retrieval without reasoning: Models are forced to answer using only the initial chunks provided, even if they are incomplete or irrelevant, without the ability to request more context.
Chunk-level limits: Critical data in multi-modal documents (such as tables or footnotes) is often lost when a model cannot open a full document or navigate to a specific region.
Lack of iteration: One-shot RAG cannot refine search queries, follow references between documents, or compare multiple sources to verify an answer.

Technical Implementation: The Mistral Search Toolkit

Agentic Search is powered by the Mistral Search Toolkit, which provides open modules for ingesting, embedding, and indexing data. The system grants the AI model five specific tools that function like file-system operations:

search: Finds relevant documents across the corpus using the existing index.
open: Opens a specific document.
navigate: Moves to a specific page, section, or region within a document.
read: Retrieves the content at the current location.
grep: Finds a specific pattern within an open document.

This architecture ensures that retrieval quality scales with the model's reasoning capabilities rather than being limited by a fixed chunking strategy, as the tools do not require model-specific fine-tuning.

Performance Benchmarks

Mistral tested Agentic Search using Mistral Medium 3.5 (MM 3.5) and Z.ai GLM-5.2 (GLM-5.2) across two industry-standard benchmarks.

FinanceBench (SEC Filings)

Testing on 368 SEC filings (~53,900 pages), the results showed:

Accuracy Gains: Moving from one-shot RAG to a search-only agentic loop increased accuracy by +47.3 percentage points (pp) for MM 3.5 and +52.6pp for GLM-5.2, representing a ~3x improvement in correctness (from 26.7% to 86%).
Navigation Impact: Adding the open, navigate, read, and grep tools further increased accuracy by +8.7pp for MM 3.5 and +6.7pp for GLM-5.2.
Efficiency: The full loop reduced token usage by 23.9% for MM 3.5 and 33.7% for GLM-5.2 compared to the search-only loop.
Latency: p90 latency dropped from 255s to 154s, and mean latency decreased from 108s to 71s.

OfficeQA Pro (Treasury Bulletins)

Testing on 696 scanned, table-heavy government-finance PDFs (~89,000 pages) showed:

Accuracy Gains: The full agentic loop increased accuracy to 51.9% for GLM-5.2 (+45.6pp) and improved accuracy by +27.1pp for MM 3.5 compared to one-shot RAG.
Efficiency: Navigation tools reduced token consumption and decreased turns by up to 7.0%.
Harness Comparison: GLM-5.2 scored 51.9% on the Mistral harness, compared to 41.4% when using the Claude Code harness, a +10.5pp difference.

Use Case Applicability

Mistral distinguishes between when to use Agentic Search versus traditional indexed retrieval:

Use Agentic Search for:

Long, dense documents: Filings, contracts, and technical specifications where answers reside in specific tables, clauses, or footnotes.
Cross-source research: Tasks requiring the reconciliation of evidence from multiple documents.
Verifiable answers: Financial or legal data where the response must be linked to a stable document location.
Structured data: Scanned PDFs and financial statements where meaning depends on spatial layout (rows/columns).

Use Indexed Retrieval (One-Shot RAG) for:

Direct lookups: Short documents where the answer is likely in the first few retrieved chunks.
High-volume search: Simple keyword or semantic lookups that do not require reasoning.
Predictable questions: Use cases where the source and location of the answer are known in advance.

Availability and Integration

Agentic Search is available through the Mistral Search Toolkit for custom agent integration and is built into Libraries within Mistral Studio and Vibe. Users can begin testing via the Search Starter App, which provides a local index for custom corpora.