RAG Document Updates

Task

A customer updated a device guide, but the assistant still quotes the previous revision. Follow a request through the supplied repository, reproduce the stale answer, and fix the document update path. The retrieval adapter returns a quoted passage so that every check is repeatable; you do not need an embedding model or an API key.

Requirements

1. run(documents, actions) starts a fresh store with generation "initial". Process actions in order and return {active_generation, results}, one result per action. Do not change either input.
2. A query action is {op: "query", text}. Return {generation, answer, sources}. Rank documents by the number of distinct shared lowercase ASCII alphanumeric words; ties use document id ascending. Return the best matching document's exact text and [{document_id, revision}]. No match returns answer null and sources [].
3. A publish action is {op: "publish", generation, documents}. It replaces the whole corpus and returns {op: "publish", generation, status: "published"}. Every subsequent query must use that generation, including previously cached queries and previously missing answers. An empty corpus removes all documents.
4. Document ids must be unique non-blank strings; revision is a positive integer, not a boolean; text is a non-blank string. Raise Error containing document for invalid documents. Generations are new non-blank strings and may not overwrite an existing generation; invalid values raise Error containing generation. Invalid query text raises Error containing query.
5. Inspect and edit the relevant modules, not only answers/service.py. Browser checks import the edited Python repository together. Public assessment checks are fixed independently of editable test files. See docs/requests.md for examples and boundaries.