Repository navigation
9.4 Embed promises (2 pts) #60
Description
Epic: E9 — Promise collection and extraction
Points: 2 · Priority: Must · Iteration: 6
As the system, I want each promise embedded, so that it can be matched to bills.
Acceptance criteria:
Every promise has a vector_id; its metadata includes promise_id and person_id.
ZeroDB data: vectors: promises
Source: backlog.md, story 9.4
Activity
developer-ainative commentedon Oct 20, 2026
Closed by PR #182 (merged). Important finding attached — read before assuming real promises exist to embed.
app/ingest/promise_embedding.pycloses the two hooks story 9.3 explicitly left open (vector_id: Noneand no event). 100% coverage, 47 tests, full suite green.The real extractor is still fabricating promises attributed to real sitting members — now with fresh, worse evidence than story 9.3's original finding
Before writing anything, this story ran the real extraction pipeline for real against all 6 live
promise_sourcesrows: 264 passages → 8 'commitments', 0 testable, all 8 visibly false positives — worse than 9.3's original 0-for-7. Specifics:electjon.comnow 302-redirects to an ActBlue donation page, so two of the eight are fundraising legal boilerplate read as a campaign promise; two are a Pelosi sentence criticizing an opponent misread as her own commitment; two are a Collins biography line. These are fabricated promises attributed to three real named public figures, including two sitting members of Congress.None of this was embedded. The live subject used for verification is
person:TEST-EMBED-9-4/ 'SYNTHETIC TEST CANDIDATE (not a real person)' with a made-up test quote, following story 6.2's established synthetic-labeling pattern — deleted afterward with a live re-read confirming the realpromisestable is empty, exactly as it was before this story ran.This is the second independent confirmation that story 13.3's standing flag is correct: story 13.2's real Q&A agent, and by extension any feature that presents extracted promises as real, must not ship without a real semantic extractor replacing the current rule-based stub. The gap isn't theoretical — it just produced 8 concrete fabrications against real people in this very verification run.
Also corrected: the live project has 6
promise_sources(allcampaign_website), not 11 across 4 types as story 9.2's own PR described — 9.2's own verify script correctly cleans up after itself on every run, so its archived rows don't persist. Not a bug, a clarification of what's actually live right now.Real live API finding: ZeroDB's
embed_and_storereturns no vector id in its response body (confirmed against the live response shape:{success, vectors_stored, embeddings_generated, model, dimensions, ...}) — so getting the realvector_idback requires a second list-and-match read, now built and tested against the ~60 req/min rate ceiling.