Agentic RAG for fintech: paying latency for precision

The failure modes that break retrieval on financial corpora are rarely exotic. They are acronyms, compound questions, and queries phrased nothing like the documents. This paper builds a four-agent RAG pipeline for an enterprise fintech knowledge base that attacks exactly those, then reports the bill. Every component addresses a failure a practitioner will recognize from the first week of running RAG on financial documents.

The four agents divide the work. A query-reformulation agent rewrites the user’s question into retrieval-friendly form. A decomposition agent splits compound questions into iterative sub-queries, guided by keyphrase extraction; a question about two products in three jurisdictions stops being one impossible embedding lookup. An acronym-resolution agent expands domain abbreviations in context, the quiet killer in fintech corpora where the same three letters mean different things on different desks. A cross-encoder reranker then reorders the retrieved candidates by reading query and passage together.

A concrete diligence question shows what the pipeline buys. Ask a naive RAG system for the change-of-control provisions across the latest credit agreements in a data room and a single embedding lookup returns boilerplate, since every agreement contains near-identical defined-terms sections. The agentic pipeline splits the question per document, expands the defined terms, then reranks evidence within each sub-question.

The evaluation is small and honestly framed: 85 question-answer-reference triples curated from the enterprise knowledge base, against which the agentic system beats baseline RAG on retrieval precision and relevance while conceding increased latency. No miracle numbers, no leaderboard sweep. The result reads less like a breakthrough and more like an engineering invoice: precision costs hops, and hops cost time.

That invoice is the useful part. The components here are the agentic versions of moves this archive has tracked all year: FinSage got its compliance-QA gains from metadata, a tuned reranker, and hybrid retrieval rather than a bigger model. The acronym agent is new in emphasis and overdue in practice. Financial text is the densest acronym environment outside the military. An embedding of “ACH return rate by SKU” retrieves garbage if the expansion lives only in a human’s head. Resolving abbreviations before retrieval is cheap, model-agnostic, and probably the highest win-per-line-of-code item in the whole pipeline.

The pipeline reads differently across the industry, because each segment hits a different failure mode first. An investment-banking deal team lives in compound questions over data rooms: covenant terms across five credit agreements, precedent clauses across a sector, exactly the multi-document queries decomposition exists for, with hours of latency tolerance and zero tolerance for a missed clause. A private-equity team runs diligence and portfolio monitoring over the least standardized corpus in finance, CIMs, quality-of-earnings reports, management presentations, where every portfolio company speaks its own dialect and term resolution does the heavy lifting. A hedge-fund desk queries filings and transcripts at scale, where the marginal query must stay cheap, making the routing tier the load-bearing component: screens take the fast path while deep dives earn the agents.

The components stay identical across segments. The tuning, latency budgets, and failure tolerances do not.

The latency concession deserves respect rather than a wince. Each agent is an LLM call; a four-stage pipeline turns one retrieval into a half-dozen model invocations, fine for an analyst’s research question, wrong for anything interactive at scale. The architecture answer is routing by query difficulty: simple lookups skip the agents entirely, while compound or acronym-heavy queries earn the full treatment. That mirrors how Memory-R1 treats memory operations, as decisions worth spending intelligence on only when the cheap path fails. A pipeline that runs every agent on every query is paying the maximum toll for traffic that mostly did not need the bridge.

The 85-triple evaluation is this note’s caveat and its lesson in one. It is far too small to generalize from. It is also exactly what a real fintech team’s first eval looks like: dozens of questions, curated from actual usage, with reference answers checked by people who know the corpus. Build that before adding any agent, because without it the latency-precision trade is a matter of opinion. The right reading is a menu rather than a verdict: four interventions, each separately testable against your own failure log. Start with acronym resolution and the reranker, measure, then let the failure cases argue for decomposition. Building the triple set is the real first step: harvest questions analysts actually asked, attach the reference passage a human judged correct, then tag every failure by cause, wrong document, right document wrong span, unresolved acronym, compound question. The taxonomy tells you which agent to add. The triples tell you whether it worked. Eighty-five curated examples beat eight hundred synthetic ones for this job, because the failure distribution is the asset.

Agentic RAG for financial corpora is an itemized bill: acronym resolution, decomposition, and reranking each buy precision at a latency price, so route the hard queries to the full pipeline and let the easy ones skip it.