From c64104e90e323ba685e54b8bf425ba7f57602df5 Mon Sep 17 00:00:00 2001 From: funsaized Date: Thu, 13 Aug 2026 18:00:28 -0400 Subject: [PATCH] docs(state): record the HUB-043 deploy verification Rebuilt research-hub and research-worker from 63aacc6, SHA-verified the six changed modules against both running containers, and exercised the corpus-wide path on the deployed API: an unfiltered /query drawing on four sources across different jobs with dual-channel fusion, a topic_filter narrowing to three on-topic sources, an unmatched filter returning nothing rather than widening, and /rag answering with citations over four distinct sources. Persisted state is unchanged: attempt-11 report and source registry still 068d60b2/d6748d76, v4 seal 762e7a19, 68,072 Qdrant points, 679 documents. Co-Authored-By: Claude Opus 5 (1M context) --- docs/CURRENT_STATE.md | 22 ++++++++++++++++++++++ 1 file changed, 22 insertions(+) diff --git a/docs/CURRENT_STATE.md b/docs/CURRENT_STATE.md index 450ba4a..7ba24e4 100644 --- a/docs/CURRENT_STATE.md +++ b/docs/CURRENT_STATE.md @@ -57,6 +57,28 @@ a `document_id` and every one resolves in SQLite, so corpus-wide retrieval drops nothing. (33 retained documents have no Qdrant chunks — deduplicated sources — and remain lexically reachable.) +**Deployed and verified on the stack.** `research-hub` and `research-worker` +rebuilt from `63aacc6`; `app/{retrieval,query,document_store,clients,main, +observability}.py` SHA-256 verified identical between the repo and both +running containers. `/readyz` reports `capability: all` with all six services +true. + +Live on the deployed API: + +| Call | Result | +|---|---| +| `/query` "how does reciprocal rank fusion combine rankings" | 6 chunks from **4 sources** across different jobs; top 3 dual-channel `dense,lexical` | +| `/query` + `topic_filter` "Kubernetes pod autoscaling…" | 4 chunks from 3 sources, all on-topic | +| `/query` + an unmatched `topic_filter` | **0 chunks** — the filter narrows, it does not silently widen | +| `/rag` "microservices versus monoliths tradeoffs" | answered with citations over 4 distinct sources | + +Persisted state after the deploy is unchanged: the attempt-11 report and +source registry are still byte-identical (`068d60b2…`, `d6748d76…`), the v4 +seal is `762e7a19…`, Qdrant holds 68,072 points, and the store holds 679 +documents / 758 observations / 71,125 lexical rows. `hub_retrieval_duration_seconds` +is live and `hub_retrieval_score` recorded 18 observations across the runs +above. + ## Live report generated through the deployed judge gate (2026-08-13) The judge pivot had been proven structurally but no report had ever been