Mistral Agentic Search lifts FinanceBench accuracy to 86%

On August 20, 2026, Mistral AI released Agentic Search, an interactive retrieval and verification layer that raises FinanceBench accuracy from 26.7% to 86%. For developers building enterprise RAG systems, the release shifts retrieval from a single lookup into a multi-step investigation.

Retrieval as an active loop

Traditional RAG systems retrieve a fixed set of text chunks and pass them to a model for an immediate answer. This approach can miss information in tables, footnotes, page references, and related filings, even when the relevant document is present in the index.

Agentic Search gives the model control over retrieval decisions. It can search for initial sources, open specific documents, navigate to relevant pages or sections, read contiguous ranges, and grep for exact terms before producing an answer. The model can repeat this process when the first retrieval pass leaves gaps.

Tool	Retrieval role
search	Finds relevant sources in indexed collections
open	Opens a target file or document
navigate	Moves to a page, section, or document location
read	Retrieves contiguous content across a range
grep	Matches exact terms inside an open file

Corpus-level pagination, including parameters such as exclude_ids, helps the agent explore new chunks instead of repeatedly receiving the same results. The system works without model-specific fine-tuning, which lets teams apply the approach to existing search infrastructure.

Benchmark results

The largest reported gain appears on FinanceBench, where questions require evidence gathered across complex SEC filings. On Mistral Medium 3.5, moving from single-shot RAG to an iterative search loop improved accuracy by 47.3 percentage points. The corresponding improvement on GLM-5.2 was 52.6 percentage points.

Evaluation	Baseline	Agentic Search result	Change
FinanceBench overall	26.7%	86%	Up to 3x correctness
OfficeQA Pro	6.3%	51.9%	+45.6 percentage points
p90 latency	Baseline	Reduced	Up to 39.6%
Token usage	Baseline	Reduced	23.9% on Mistral Medium 3.5, 33.7% on GLM-5.2

Active navigation contributed an additional 8.7 percentage points on Mistral Medium 3.5 after the iterative loop was introduced. OfficeQA Pro produced a similar pattern on scanned, table-heavy Treasury Bulletin documents, where accuracy increased by 45.6 percentage points.

The efficiency results are significant because deeper retrieval commonly increases model calls and context volume. Targeted section reading lets the agent spend retrieval steps on likely evidence rather than repeatedly re-indexing or passing broad document ranges. Mistral reports p90 latency reductions of up to 39.6% and token reductions of up to one-third.

Deployment paths

Agentic Search is available through the Mistral Search Toolkit as an SDK and composable framework for production retrieval pipelines. It is also integrated into Libraries in Mistral Studio and Vibe, and can run as a set of Model Context Protocol tools in an existing agent workflow. Teams working with MCP can therefore add the retrieval primitives without replacing their broader orchestration layer.

Deployment can span cloud, on-premises, and sovereign infrastructure. Enterprise data can remain within the organization’s chosen environment, an important constraint for financial filings, regulated records, and internal document collections.

Implications for RAG systems

The architecture moves more responsibility from the retriever into the agent. Retrieval quality becomes partly a tool-use problem: the model must decide when to inspect a document, which page to open, how much content to read, and whether exact-term verification is necessary.

If you maintain a production RAG pipeline, evaluate multi-hop and table-heavy questions separately from ordinary semantic retrieval. Track answer correctness alongside tool-call count, p90 latency, and tokens consumed. Agentic Search is most relevant when evidence is distributed across long documents, while simple fact lookup can continue using a lower-cost single-pass path.

The practical next step is to test the five primitives against your hardest enterprise queries, preserve document-level provenance for every read operation, and set budgets for navigation depth, latency, and token usage before moving the workflow into production.