Retrieval That Lawyers Would Actually Trust

The repository connector captured a stable document identifier, version, source location, content type, access attributes and lifecycle state before extraction. Parsers produced structured blocks rather than one undifferentiated string, retaining headings, list scope, page anchors and table relationships where the format allowed. Optical character recognition outputs carried confidence and were never silently substituted for a failed native extraction. Quality checks identified empty pages, repeated headers, suspicious character sequences and abrupt loss of content. A document with an ingestion failure remained visibly absent from search; it was not marked indexed merely because a processing job had completed.

Chunks followed document structure before applying a token ceiling. Definitions stayed with their terms, headings with their subordinate text and table rows with the headers needed to interpret them. Each chunk retained offsets into the immutable extracted version, permitting a citation to resolve to the exact evidence shown during answer generation. Re-indexing created a new index version and preserved the previous evaluation result until comparison completed. Deletes and permission changes propagated through lexical, vector and cache layers. This lifecycle prevented an old vector from surviving after its source had been restricted or replaced, a failure that answer-time instructions could not safely repair.

A transparent lexical baseline exposed questions driven by exact terms, authorities and clause references. Dense retrieval added semantic candidates for questions phrased differently from the source, but it did not replace term search. Candidate sets were fused and passed to a cross-encoder reranker, with document status, effective date and source class available as explicit signals or filters. The retrieval stage favoured evidence coverage; context assembly later controlled volume. Tests were sliced by query intent and document type because a single average could conceal consistent failure on tables, defined terms or superseded policies.

Permission filters were incorporated into each search plan. Applying them after nearest-neighbour retrieval could leave an authorised user with poor candidates even where relevant permitted evidence existed, so filter behaviour was measured against realistic matter scopes. Query traces recorded normalisation, candidate identifiers, scores, filters, index version and selected chunks without copying unrestricted content into general telemetry. Failure analysis separated absent source, extraction damage, boundary error, vocabulary mismatch, permission error and ranking error. Engineers could then repair the earliest failing stage instead of compensating with a larger model or a prompt that merely sounded more cautious.

Top-ranked passages frequently overlapped or repeated the same source. Context assembly grouped adjacent material where continuity improved meaning, removed near-duplicates and limited dominance by one document when independent sources were relevant. It preserved the heading path, document status and citation anchor alongside each passage. The token budget was allocated to evidence rather than repository boilerplate. Contradictory current passages were retained and flagged to the answer layer rather than arbitrarily selecting one. Superseded material could support a historical question, but its status travelled into the context and display so it could not appear as present guidance by omission.

The answer contract required claims to be supported by supplied passages and separated quoted source meaning from synthesis. Where labelled evidence was absent or conflicting, abstention was an accepted outcome. Citation verification checked that the referenced passage entailed the nearby claim, not simply that it came from a relevant document. Stable identifiers resolved citations through the user’s current access, and the display named the exact revision used. If access had since changed, the historical trace remained in controlled logs without disclosing the content. Generation settings, prompt version, chosen chunks and model version were recorded so a disputed answer could be replayed against the same inputs.

The evaluation corpus paired each question with the smallest passages sufficient to answer it and recorded acceptable alternatives. It also contained similarly named matters, obsolete guidance, missing-source questions and requests that crossed access boundaries. Retrieval was measured independently through evidence coverage and ordering, while reviewers assessed answer support, citation accuracy and appropriate abstention. A model answer that happened to be correct from prior knowledge did not rescue failed retrieval. Likewise, stylistic disagreement did not count as a factual failure where the response remained grounded and complete. This separation made changes to embeddings, parsers and prompts comparable rather than anecdotal.

Production feedback entered a triage process rather than an automatic training loop. Reviewers could identify a wrong source, damaged extraction, missed qualification, unsupported claim or access concern against the recorded trace. Approved cases joined the regression set after sensitive details were handled under the same corpus controls. Index freshness, deletion propagation, parsing failures and permission-sync lag were monitored as product health, because retrieval quality degrades as sources change. The delivered system therefore made trust a property of inspectable evidence and controlled failure. It did not ask lawyers to trust the language model, and it did not present an eloquent answer as proof that research had succeeded.