christian/grant-outreach-engine
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity
Files
cbc4512ffaeb09675b2bfaa601a73da8c68a2d2a
grant-outreach-engine/docs/features/profiler.md
Croissant Le Doux 4f35dcdb8d feat(profiler): Stage 4 grounded research profiler — cited profiles replace NTEE stubs 
profileOrgs (nightly 05:45, 25/run, easy-win orgs best-first per the
plan's coarse-match-gates-profiling rule): two-pass Gemini flash —
Google Search-grounded research (groundingMetadata = citation universe)
then JSON-schema extraction citing only from it. Overwrites the stub
profile (mission/programs/geography/funders/staff/news/sources/
confidence); re-embeds mission fit from researched text unless identity
unconfirmed or confidence <0.5 (then the NTEE embedding stays —
unverified research must not steer scoring). Detail page renders
researched programs w/ source links, funders, staff.

Fixes: org_profiles.confidence is float4 — 0.2 stores as 0.20000000298
so 'confidence <= 0.2' excluded every stub (epsilon comparison); RR7
loader serialization turns unknown into never (hoisted casts).

Live: 55 orgs profiled (41 research-grade, 2 low-confidence, 0 failed).
Scoring effect verified both directions: Annie's Angels x NIH mammalian
models fit 19->12; Seacoast Pathways x Foundation for Seacoast Health
fit 25/30 — and the researched knownFunders independently names that
funder. 161 tests green.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-17 00:08:12 -04:00

2.1 KiB
Raw Blame History

Stage 4 — Org research profiler

Nightly profileOrgs workflow (05:45, after matchGrants): grounded web research for orgs that already have ≥1 non-rejected easy-win match (the plan's "coarse match gates profiling" cost rule), best-scoring first, 25/night (PROFILE_ORGS_PER_RUN).

Two Gemini passes per org (packages/outreach-ai/src/agents/org-profiler/research.ts):

  1. Grounded research — gemini-2.5-flash with the Google Search tool; groundingMetadata supplies the citation universe (every page actually consulted).
  2. Structured extraction — flash constrained to a JSON schema over the research notes, citing only from that universe. Search grounding and JSON-schema output can't share one call, hence the split.

Output overwrites the org's NTEE stub in org_profiles (mission, programs w/ populations served, service geography, known funders, staff, recent news, sources, confidence). The profile embedding is rebuilt from the researched text — unless identity wasn't confirmed or confidence < 0.5, in which case the embedding stays NTEE-derived so unverified research can't steer mission fit. An identity-unconfirmed profile is stored at ≤0.3 confidence: "we looked and couldn't confirm" beats an NTEE guess, and the review UI warns on it.

Review integration: match detail pages render researched programs (with per-claim source links), known funders, and staff.

First live batches (2026-07-16/17): 55 orgs, 41 research-grade (>0.4), 2 low-confidence, 0 failures, ~28 min for 30 orgs (grounded search dominates). Verified effect on scoring: Annie's Angels → NIH Mammalian Models mission fit fell 19→12 (real mission: family financial crisis support), Seacoast Pathways → Foundation for Seacoast Health rose to 25/30 fit — and its researched knownFunders list independently named Foundation for Seacoast Health, confirming the precedent match.

Gotcha fixed en route: org_profiles.confidence is float4 — 0.2 stores as 0.20000000298, so confidence <= 0.2 (float8 comparison) silently excluded every stub; the needing-profile query uses an epsilon.