EPIC Search Business Rules

EPIC.search business rules

EPIC.search is a separate EAO search product: a React web app, two Flask services (search-api for orchestration and model calls, search-vector-api for retrieval) and a Postgres/pgvector embedder. Its source is the EPIC.search repository; a workspace clone goes in EPIC.search/ beside eagle-demi.

None of its rules govern DEMI. Nothing in eagle-demi/src references the product, and no rule read out of it has been promoted into DEMI policy. DEMI's rules are Business-Rules (BR-01 to BR-09), its retrieval design is Search-and-Retrieval, and its access model is ADR-004-Read-ACL-Authorization-Model with Identity-Model.

Where DEMI decides differently

Topic	EPIC.search	DEMI
Retrieval	Vector search over pgvector, re-ranked by a cross-encoder.	Lexical BM25 in Azure AI Search, no embeddings — ADR-005-Cosmos-Full-Text-Search-Ruled-Out, ADR-006-AI-Summarization-over-BM25.
Role of the model	The model gates query scope, classifies query complexity, infers project and document-type filters, and rewrites the query before retrieval.	The model summarises chunks the lexical query already returned. It never selects documents, re-ranks or rewrites (src/ai/summarize.js).
Summary grounding	The summariser prompts carry no grounding, citation or no-fabrication rule.	The summariser answers only from the numbered sources, cites every claim, and says so when the sources do not answer (src/ai/summarize.js).
Public access	Every search needs a BC Gov IDIR login; there is no anonymous route.	Anonymous search is served, and the caller's read[] decides what it returns (src/http/routes.js).
Published gating	document_status is stored and indexed at ingest but never used as a search filter.	A document is public only when it and its parent project are both public — BR-04.
Metadata classification	The embedder tags every chunk against a fixed 111-term vocabulary, by literal or semantic match.	Reuse of that embedder is rejected in favour of rule-based suggestions a person confirms — ADR-003-Rule-Based-Classification.
Ingest admission	pdf, images, docx and plain text only; doc, xlsx and pptx are refused.	pdf, doc, docx, pptx and xlsx are all extracted (src/extract.js:32-38).
Chunking	1000 characters with 200 overlap, chunked one page at a time.	Sections accumulate to a target size and overlap across block boundaries; pageNumber is a sequence number, not a PDF page (src/chunker.js).