01
Evidence before generation
Claude receives a deliberately assembled evidence pack, not unrestricted access to the entire corpus.
01 / The system
Research, not guesswork.
The product treats every legal answer as an evidence pipeline rather than a single model call.
Global regulation is fragmented by jurisdiction, regulator, document type, and time. The RAG makes that complexity searchable.
The platform ingests regulatory material from multiple markets, preserves the legal metadata that gives each passage meaning, and retrieves the right evidence before Claude writes an answer.
The key engineering decision was to make jurisdiction, regulator, instrument, publication date, and effective date first-class retrieval fields. That prevents a semantically similar rule from the wrong market or the wrong point in time from silently contaminating an answer.
02
Legal-aware retrieval
Semantic similarity is combined with exact jurisdiction filters, lexical signals, and re-ranking.
03
Traceable outputs
Every material claim is designed to map back to a source, passage, and document version.
02 / Architecture
The evidence pipeline.
Six focused stages turn a shifting regulatory corpus into reviewable intelligence.
World regulations RAG / System map
Citation-first path
Stage 01
Source acquisition
Collect regulatory pages, circulars, consultation papers, guidance, and PDFs from authoritative sources.
Connectors · Jobs · VersioningStage 02
Parse & normalize
Extract clean text, retain headings and tables, remove navigation noise, and standardize document structure.
Python · Parsers · OCR fallbackStage 03
Legal chunking
Split by article, section, and semantic boundary while inheriting jurisdiction and document metadata.
Structure · Overlap · MetadataStage 04
Hybrid retrieval
Combine dense vector search, keyword relevance, metadata constraints, and document diversity.
Embeddings · BM25 · FiltersStage 05
Claude synthesis
Build a bounded evidence pack and ask Claude to compare, qualify, cite, and abstain when evidence is weak.
Claude API · Structured outputStage 06
Answer & audit
Return structured conclusions, inline citations, source excerpts, and review metadata through the API.
FastAPI · Schemas · LogsThe model is one layer, not the system.
Retrieval, filtering, context construction, validation, and observability do most of the reliability work around the LLM.
Every stage is debuggable.
A weak answer can be traced to ingestion, metadata, retrieval, ranking, prompting, or citation mapping instead of being treated as an opaque failure.
03 / API engineering
Claude behind a contract.
A typed Python service turns model capability into a predictable application interface.
POST /v1/regulatory/research
# Python client request payload = { "question": "Compare marketing controls", "jurisdictions": ["AE", "EU", "SG"], "as_of": "2026-07-01", "answer_mode": "comparative", "citations": True } result = client.post( "/v1/regulatory/research", json=payload, timeout=60 ) # Validated response schema { "answer": "...", "jurisdiction_findings": [...], "citations": [...], "review_flags": [] }
Typed inputs
Pydantic validation rejects ambiguous jurisdiction codes, invalid dates, and unsupported answer modes before retrieval begins.
Bounded model calls
The service sends Claude only ranked evidence and explicit output rules, with timeouts, retry policy, and token budgets controlled in Python.
Structured responses
Applications receive stable fields for findings, sources, confidence signals, and review flags instead of a blob of unpredictable prose.
Operational visibility
Request IDs connect API latency, retrieval traces, model usage, citations, and failures across the full execution path.
04 / Retrieval
Similarity is not enough.
Legal relevance depends on where, when, who, and under which instrument a passage applies.
Metadata narrows the legal universe before ranking begins.
The retriever first constrains the corpus with deterministic legal context. Dense and lexical results are then fused, re-ranked, and diversified so one long document cannot crowd out every other authority.
- JurisdictionExact / regional scope
- RegulatorIssuing authority
- InstrumentRule · guidance · consultation
- TimePublished · effective · superseded
- DomainLicensing · AML · marketing
Illustrative retrieval trace
Fusion + re-rank01Marketing communication requirements
02Crypto-asset communications framework
03Consumer-protection expectations
04Risk statement and disclosure controls
05Cross-border promotion considerations
Deterministic filtersDense + keywordDiversity pass
05 / Trust layer
Designed to know its limits.
The answer contract makes uncertainty and evidence quality visible to the reviewer.
Control 01
Citation coverage
Material claims must point to retrieved passages. Unsupported statements are removed or marked for review.
Claim → sourceControl 02
Temporal awareness
Effective dates and superseded versions stay attached to chunks so historical rules do not masquerade as current.
As-of dateControl 03
Abstention path
If evidence is thin, conflicting, or outside the selected scope, the system says so instead of completing the gap.
No-answer modeControl 04
Human review
Source excerpts, document links, and review flags keep a qualified person in control of the final interpretation.
Decision support06 / Evaluation
Quality at every layer.
RAG evaluation separates retrieval failures from answer failures so improvements are targeted.
Retrieval evaluation
A curated set of regulatory questions tests whether the correct authority and passage reach the context window across markets and document types.
Relevant passage found in top resultsRecall@K
Correct authority ranked earlyMRR
Results cover required jurisdictionsCoverage
Superseded documents excludedTemporal
Answer evaluation
Generated answers are reviewed against the retrieved evidence and the actual research intent. Fluency alone is not the score.
GroundingEvidence check
Does every conclusion follow from a cited passage?
CompletenessScope check
Were all requested jurisdictions and sub-questions addressed?
Citation precisionAttribution
Does each citation support the exact sentence it follows?
Calibrated uncertaintySafety
Does the answer qualify conflict, ambiguity, and missing evidence?