Retrieval-Augmented Generation

Contract and corpus

Define users/questions, answer/citation requirements, freshness, latency/cost, permissions, deletion, languages, and failure policy. Inventory source ownership, formats, versioning, sensitive data, access controls, and update cadence. Do not index content the application is not authorized to reveal.

Pipeline

1. Parse deterministically and preserve source ID, version/hash, page/section/line, title, ACL, and timestamps.
2. Chunk on semantic/structural boundaries with limited overlap; keep tables/code and parent context intentionally.
3. Build a lexical baseline. Add embeddings only when evaluation shows benefit; keep backend/model configurable and record exact versions.
4. Filter by authorization before retrieval and again before answer assembly.
5. Retrieve bounded candidates, combine lexical/semantic ranks when useful, rerank a small set, diversify duplicates, and pass only relevant evidence.
6. Instruct the answer layer to use provided evidence, cite source IDs/ranges, distinguish inference, and abstain when support is insufficient.
7. Incrementally update by content hash and remove stale/deleted chunks and caches.

External embedding/vector/reranking services require authorization and a privacy/retention review. Do not install infrastructure or send private corpora outside the approved boundary automatically.

Evaluation

Create labeled queries with relevant sources and expected abstentions. Measure retrieval recall@k/MRR or nDCG, citation correctness, answer support, unsupported-claim rate, ACL leakage, freshness, latency, cost, and failure recovery. Compare no-retrieval, lexical, semantic, and hybrid baselines. Test injection inside documents, conflicting sources, stale updates, deleted documents, and permission changes.

Completion report

Document corpus/provenance, parser/chunker, index and versioning, ACL/deletion, retrieval/reranking, prompt/output contract, evaluation results, observability, deployment/rollback, and unresolved coverage.