A corpus you cannot ask a question of is only a warehouse. This week I add semantic search over the governed chunk table, hybrid fusion with lexical hits, ACL filtering, and hierarchical container scoping so a query can stay inside an OTCS inheritance subtree. I also add a generalized repo ingester that writes those scoped embeddings. The claim is not that recall is now excellent. The claim is that search finally has a container, a permission filter, and a place to write chunks that belong to a project.
The wrong default
The usual search is global. Everything in the vector table is a candidate. That is convenient in a demo and fatal in a records system. A question about one department’s file plan should not retrieve another department’s sentences because they sound alike. Permissions that exist on the file must still constrain the hit. A subtree in OpenText is a real container. Search that cannot see it is browsing a dump.
A map that shows every street in the country when you asked for one building is not navigation. I want the building, the floor, and the door that is allowed.
What I actually added
Semantic corpus search now runs over chunk_embeddings_v2. Hybrid RRF fuses that with lexical retrieval. ACL filtering belongs on the same path, not as an afterthought in the UI. Hierarchical container scoping uses OTCS-inheritance subtree keys so a caller can restrict the corpus the way the content platform already restricts browsing. The repo ingester writes into that same scoped table instead of a private side index. An embedding backfill exists for documents that already sit under Alfresco governance.
I also add a retrieval CLI and a corpus Q&A skill so I can ask the index from a working tree. Search telemetry and a metrics route went in and then came back out on an operator hold. I am not treating the reverted stream as a delivered observatory. What remains is the search surface itself.
Thought lifecycle sits on the same hybrid path. A thought is not a document chunk, but it has to survive the same permission and container instincts or I will have two search products that lie in different ways. The repo ingester is generalized so I am not maintaining a StoryTeller-only writer. Scoped chunk_embeddings_v2 is the table. If a document is already under Alfresco, backfill writes there rather than inventing a third index.
This is search over my own corpora. Identifying a pilot client in delivery documents the same week is not this claim. I am not turning this post into a customer name.
CI now has a full suite that can run against a Postgres service, and I made the workflow manual-only so a noisy automatic run does not pretend to be a gate. A unique index on document_metadata.file_id stops duplicate document rows. The first-deposit version-label defect from last month gets a follow-up: defer cm:versionable so the first landing gets the correct label. Those are plumbing. They are how the search path stays honest.
The next test
The next test is a query I can point at one subtree, with ACL on, and show that a similar sentence outside that container does not come back. Until that is measured, this is a search contract I can call, not a claim that the corpus already answers.
Historical basis: VALESKA commits 5925b17, c92b4e8, d64ae8b, bf5d139, and related 2–5 July 2026 work. The reverted telemetry stream is not treated as delivered.