Elicon
← ALL REFERENCE ENGAGEMENTS
REFERENCE ENGAGEMENT
Capital Markets & Brokerage

Compliance Search for a Regulated Financial Firm

Hybrid search over every regulation and policy, with cited, approvable audit answers.

  • Python
  • FastAPI
  • pgvector
  • Hybrid search
  • Postgres

Reference engagements illustrate how Elicon would approach a project of this kind — the problem, the architecture and the decisions. They are examples, not accounts of past client work.

THE SITUATION

The problem.

A regulated financial firm whose compliance officers answer audit questions by searching PDFs one at a time, with answers that vary by officer. The engagement builds hybrid semantic search over every regulatory and internal policy document, a drafting interface where the citation trail is the deliverable, and a library of approved answers.

Audit responses take days of manual searching and copying, and two officers can give different answers to the same question.

  • ·Keyword search misses paraphrased questions; meaning-based search misses control codes and regulation identifiers.
  • ·Previous answers are not reusable, so every audit starts from scratch.

APPROACH

How we'd run it.

  1. 01Ingest every regulatory and policy document with OCR, section-aware chunking and version history.
  2. 02Combine keyword and vector search so both exact identifiers and paraphrased questions work.
  3. 03Draft cited answers for officers to edit and approve.
  4. 04Store approved answers with their citations so the next audit starts from the last one.

ARCHITECTURE

What gets built.

Ingestion & versioning

OCR where needed, section-aware chunking, and a version graph per document. Superseded content is retrievable only when a query asks for history.

Hybrid retrieval

BM25 and dense vectors over the same chunks, fused with reciprocal rank fusion and reranked. Control codes hit exactly; questions about intent hit semantically.

Drafting interface and answer library

The officer picks an audit question and gets a draft with a citation on every assertion, editing in place. Approved answers become retrievable themselves.

Access & logging

Single sign-on, per-document access rules mirroring existing permissions, and an append-only log of every query, retrieval and approval.

DECISIONS

The trade-offs.

Hybrid retrieval from the start rather than vectors first

Dense-only retrieval fails on control codes and keyword-only fails on paraphrase. Real audit questions contain both.

TRADE-OFFMore moving parts and a fusion step to tune.

A version graph instead of a current-only index

Auditors ask what the policy was at a past date. Deleting history to keep the index clean removes answers the firm may be required to produce.

TRADE-OFFMore ingestion complexity, and a retrieval filter enforced at the query layer so it can't be forgotten.

Draft and approve rather than auto-answer

Compliance answers carry personal accountability. Officers will not sign what they have not shaped.

TRADE-OFFLess automation on paper; in practice the drafting step is where the time goes.

STACK

Typical stack.

BACKEND
PythonFastAPIPostgres
RETRIEVAL
pgvectorBM25Reciprocal rank fusionReranking
INGESTION
Layout-aware extractionOCRVersion graphSection chunking
GOVERNANCE
SSOAccess loggingApproval workflow

Work of this shape usually runs as AI Product Launch →

SIMILAR PROBLEM?

Tell us what you're dealing with.

One conversation with an engineer is usually enough to tell whether this is a few weeks of work, a few months, or something you shouldn't build at all.

Talk to an engineer →