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>
This commit is contained in:
Croissant Le Doux
2026-07-17 00:08:12 -04:00
parent 88affbeb4f
commit 4f35dcdb8d
14 changed files with 765 additions and 0 deletions

View File

@@ -63,6 +63,17 @@ export default function MatchDetail({ loaderData }: Route.ComponentProps) {
restrictionsOnAwards?: string; restrictionsOnAwards?: string;
} | null; } | null;
const subscores = (match.subscores ?? {}) as Record<string, number>; const subscores = (match.subscores ?? {}) as Record<string, number>;
const orgPrograms = (org.programs ?? null) as Array<{
name: string;
description?: string | null;
populationServed?: string | null;
sourceUrl?: string | null;
}> | null;
const orgFunders = (org.knownFunders ?? null) as Array<{ name: string }> | null;
const orgStaff = (org.staff ?? null) as Array<{
name: string;
role?: string | null;
}> | null;
const rationale = (match.rationale ?? {}) as Record<string, unknown>; const rationale = (match.rationale ?? {}) as Record<string, unknown>;
const nhHistory = funderGivingHistory; const nhHistory = funderGivingHistory;
@@ -143,6 +154,43 @@ export default function MatchDetail({ loaderData }: Route.ComponentProps) {
</dd> </dd>
</div> </div>
</dl> </dl>
{orgPrograms != null && orgPrograms.length > 0 && (
<div className="mt-3 text-sm">
<h3 className="font-medium">Programs (researched)</h3>
<ul className="ml-4 list-disc">
{orgPrograms.map(
(p, i) => (
<li key={i}>
{p.name}
{p.populationServed != null && ` — serving ${p.populationServed}`}
{p.sourceUrl != null && (
<>
{' '}
<a className="text-blue-700 underline" href={p.sourceUrl} target="_blank" rel="noreferrer">
src
</a>
</>
)}
</li>
),
)}
</ul>
</div>
)}
{orgFunders != null && orgFunders.length > 0 && (
<p className="mt-2 text-sm">
<span className="font-medium">Known funders: </span>
{orgFunders.map((f) => f.name).join(', ')}
</p>
)}
{orgStaff != null && orgStaff.length > 0 && (
<p className="mt-2 text-sm">
<span className="font-medium">Staff: </span>
{orgStaff
.map((s) => (s.role != null ? `${s.name} (${s.role})` : s.name))
.join(', ')}
</p>
)}
<div className="mt-3 space-x-3 text-sm"> <div className="mt-3 space-x-3 text-sm">
{org.ein != null && ( {org.ein != null && (
<a <a

View File

@@ -38,6 +38,7 @@ import { setIngestGrantsDeps } from './workflows/ingest-grants.js';
import { setIngest990pfDeps } from './workflows/ingest-990pf.js'; import { setIngest990pfDeps } from './workflows/ingest-990pf.js';
import { setIngestNhdojOrgsDeps } from './workflows/ingest-nhdoj-orgs.js'; import { setIngestNhdojOrgsDeps } from './workflows/ingest-nhdoj-orgs.js';
import { setMatchGrantsDeps } from './workflows/match-grants.js'; import { setMatchGrantsDeps } from './workflows/match-grants.js';
import { setProfileOrgsDeps } from './workflows/profile-orgs.js';
import { setIngestPndRssDeps } from './workflows/ingest-pnd-rss.js'; import { setIngestPndRssDeps } from './workflows/ingest-pnd-rss.js';
if (process.env.DATABASE_URL == null) { if (process.env.DATABASE_URL == null) {
@@ -73,6 +74,7 @@ async function main() {
setMatchGrantsDeps({ db }); setMatchGrantsDeps({ db });
setIngest990pfDeps({ db }); setIngest990pfDeps({ db });
setFederalPrecedentDeps({ db }); setFederalPrecedentDeps({ db });
setProfileOrgsDeps({ db });
DBOS.setConfig({ DBOS.setConfig({
name: 'helmdocs-outreach-worker', name: 'helmdocs-outreach-worker',

View File

@@ -48,6 +48,10 @@ import {
runMatchGrantsNow, runMatchGrantsNow,
setMatchGrantsDeps, setMatchGrantsDeps,
} from './workflows/match-grants.js'; } from './workflows/match-grants.js';
import {
runProfileOrgsNow,
setProfileOrgsDeps,
} from './workflows/profile-orgs.js';
import { import {
runIngestPndRssNow, runIngestPndRssNow,
setIngestPndRssDeps, setIngestPndRssDeps,
@@ -63,6 +67,7 @@ const RUNNERS: Record<string, () => Promise<void>> = {
matchGrants: runMatchGrantsNow, matchGrants: runMatchGrantsNow,
ingest990pf: runIngest990PfNow, ingest990pf: runIngest990PfNow,
federalPrecedent: runFederalPrecedentNow, federalPrecedent: runFederalPrecedentNow,
profileOrgs: runProfileOrgsNow,
}; };
const FIRST_RUN_ORDER = [ const FIRST_RUN_ORDER = [
@@ -75,6 +80,7 @@ const FIRST_RUN_ORDER = [
'federalPrecedent', 'federalPrecedent',
'embedGrants', 'embedGrants',
'matchGrants', 'matchGrants',
'profileOrgs',
]; ];
if (process.env.DATABASE_URL == null) { if (process.env.DATABASE_URL == null) {
@@ -113,6 +119,7 @@ async function main() {
setMatchGrantsDeps({ db }); setMatchGrantsDeps({ db });
setIngest990pfDeps({ db }); setIngest990pfDeps({ db });
setFederalPrecedentDeps({ db }); setFederalPrecedentDeps({ db });
setProfileOrgsDeps({ db });
DBOS.setConfig({ DBOS.setConfig({
name: 'helmdocs-outreach-worker', name: 'helmdocs-outreach-worker',

View File

@@ -0,0 +1,208 @@
/**
* Nightly Stage 4 profiling workflow — grounded web research for orgs
* that already have a plausible lead (>=1 non-rejected easy-win match),
* best-scoring first, replacing their NTEE-stub profiles with cited,
* researched ones. Better profiles → better mission-fit embeddings on the
* NEXT match run → better personalization at review time.
*
* The nightly batch cap bounds LLM spend (two Flash calls + one embedding
* per org, ≈ cents each); env `PROFILE_ORGS_PER_RUN` overrides.
*
* Low-confidence results (identity not confirmed, or the model reports
* <= 0.4) are STILL stored — with their low confidence — so the review UI
* can warn, but they do NOT overwrite... actually they do overwrite the
* stub, which is correct: "we looked and couldn't confirm" is more
* information than an NTEE guess, and the confidence field carries it.
* The profile embedding, however, keeps mission-fit honest: identity-
* unconfirmed profiles are embedded from the stub text, not the
* possibly-wrong research.
*
* Runs at 05:45, after matchGrants (05:15). Registration follows
* `ingest-grants.ts` exactly.
*/
import { DBOS, SchedulerMode } from '@dbos-inc/dbos-sdk';
import {
buildProfileEmbeddingText,
generateQueryEmbedding,
researchOrgProfile,
type ResearchedProfileResult,
} from '@novelpad/outreach-ai';
import { buildOrgMissionText, type schema } from '@novelpad/outreach-core';
import {
serverApplyResearchedProfile,
serverListOrgsNeedingProfile,
type OrgNeedingProfile,
} from '@novelpad/outreach-core/server';
import type { NodePgDatabase } from 'drizzle-orm/node-postgres';
export type OutreachDb = NodePgDatabase<typeof schema>;
const DEFAULT_ORGS_PER_RUN = 25;
/** Below this, the research embedding is not trusted for mission fit. */
const EMBEDDING_TRUST_FLOOR = 0.5;
export interface ProfileOrgsDeps {
readonly db: OutreachDb;
}
let registeredDeps: ProfileOrgsDeps | null = null;
export function setProfileOrgsDeps(deps: ProfileOrgsDeps): void {
registeredDeps = deps;
}
function getProfileOrgsDeps(): ProfileOrgsDeps {
if (registeredDeps == null) {
throw new Error(
'ProfileOrgsDeps not registered. Call setProfileOrgsDeps() before DBOS.launch().',
);
}
return registeredDeps;
}
async function listOrgsNeedingProfile(
db: OutreachDb,
limit: number,
): Promise<OrgNeedingProfile[]> {
return serverListOrgsNeedingProfile(db, { limit });
}
const listOrgsNeedingProfileStep = DBOS.registerStep(listOrgsNeedingProfile, {
name: 'listOrgsNeedingProfile',
retriesAllowed: true,
maxAttempts: 3,
});
async function researchOrg(
org: OrgNeedingProfile,
): Promise<ResearchedProfileResult> {
return researchOrgProfile(org);
}
const researchOrgStep = DBOS.registerStep(researchOrg, {
name: 'researchOrgProfile',
retriesAllowed: true,
maxAttempts: 2,
});
async function storeProfile(
db: OutreachDb,
org: OrgNeedingProfile,
result: ResearchedProfileResult,
): Promise<void> {
const { profile, sources } = result;
const trustworthy =
profile.identityConfirmed &&
profile.overallConfidence >= EMBEDDING_TRUST_FLOOR;
const embeddingText = trustworthy
? buildProfileEmbeddingText(org, profile)
: buildOrgMissionText(org);
const embedding = await generateQueryEmbedding(embeddingText);
await serverApplyResearchedProfile(db, {
orgId: org.id,
missionStatement: profile.mission?.value ?? null,
programs: profile.programs,
serviceGeography: profile.serviceGeography?.value ?? null,
recentNews: profile.recentNews,
knownFunders: profile.knownFunders,
staff: profile.staff,
sources,
confidence: profile.identityConfirmed
? profile.overallConfidence
: Math.min(profile.overallConfidence, 0.3),
profileEmbedding: embedding,
});
}
const storeProfileStep = DBOS.registerStep(storeProfile, {
name: 'storeResearchedProfile',
retriesAllowed: true,
maxAttempts: 3,
});
async function runProfileOrgs(): Promise<void> {
const { db } = getProfileOrgsDeps();
const limit = Number(process.env.PROFILE_ORGS_PER_RUN ?? DEFAULT_ORGS_PER_RUN);
const orgs = await listOrgsNeedingProfileStep(db, limit);
if (orgs.length === 0) {
console.log('[profile-orgs] no easy-win orgs awaiting research');
return;
}
let profiled = 0;
let lowConfidence = 0;
let failed = 0;
for (const org of orgs) {
try {
const result = await researchOrgStep(org);
await storeProfileStep(db, org, result);
profiled++;
if (
!result.profile.identityConfirmed ||
result.profile.overallConfidence < EMBEDDING_TRUST_FLOOR
) {
lowConfidence++;
}
} catch (err) {
failed++;
console.error(`[profile-orgs] org "${org.name}" failed:`, err);
}
}
console.log(
`[profile-orgs] attempted=${orgs.length} profiled=${profiled} lowConfidence=${lowConfidence} failed=${failed}`,
);
if (failed > 0 && failed / orgs.length > 0.2) {
throw new Error(
`[profile-orgs] systemic failure: ${failed}/${orgs.length} orgs failed`,
);
}
}
const g = globalThis as unknown as {
__outreachProfileOrgsRegistered?: boolean;
__outreachProfileOrgsHandle?: (
scheduledTime: Date,
startedAt: Date,
) => Promise<void>;
};
if (!g.__outreachProfileOrgsRegistered) {
g.__outreachProfileOrgsRegistered = true;
const profileOrgs = async (_scheduledTime: Date, _startedAt: Date) => {
try {
await runProfileOrgs();
} catch (err) {
console.error('[profile-orgs] pass failed:', err);
throw err;
}
};
// Must be registered as BOTH a workflow and a scheduled function,
// referencing the same function object — see ingest-grants.ts.
g.__outreachProfileOrgsHandle = DBOS.registerWorkflow(profileOrgs, {
name: 'profileOrgs',
});
DBOS.registerScheduled(profileOrgs, {
crontab: '45 5 * * *',
name: 'profileOrgs',
mode: SchedulerMode.ExactlyOncePerInterval,
});
}
/**
* Starts one durable run of this workflow immediately through DBOS —
* the exact production path (workflow + checkpointed steps), used by
* `run-once.ts` for supervised/manual passes. Requires deps injected and
* `DBOS.launch()` completed.
*/
export function runProfileOrgsNow(): Promise<void> {
const handle = g.__outreachProfileOrgsHandle;
if (handle == null) {
throw new Error(
'profileOrgs is not registered; was this module imported before DBOS.launch()?',
);
}
return handle(new Date(), new Date());
}

15
docs/features/profiler.md Normal file
View File

@@ -0,0 +1,15 @@
# 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.

View File

@@ -0,0 +1,106 @@
import { getAi } from '../../gemini.js';
import { BULK_MODEL } from '../../models.js';
import {
buildExtractionPrompt,
buildResearchPrompt,
RESEARCHED_ORG_PROFILE_JSON_SCHEMA,
ResearchedOrgProfileSchema,
type ProfileResearchTarget,
type ResearchedOrgProfile,
} from './researched-profile.js';
export interface ResearchedProfileResult {
readonly profile: ResearchedOrgProfile;
/** Every grounding source the research pass touched (deduped). */
readonly sources: Array<{ url: string; title: string | null }>;
}
/**
* Stage 4 profiler: two Gemini passes.
*
* 1. Grounded research — BULK_MODEL with the Google Search tool; the
* response's groundingMetadata carries the actual pages consulted,
* which become the profile's citation universe.
* 2. Structured extraction — BULK_MODEL constrained to
* RESEARCHED_ORG_PROFILE_JSON_SCHEMA over the research text, citing
* only from that universe.
*
* Search grounding and JSON-schema output can't share one call, hence the
* split. Cost ≈ two Flash calls per org — the per-org spend the plan's
* "coarse match gates profiling" rule exists to bound.
*/
export async function researchOrgProfile(
org: ProfileResearchTarget,
): Promise<ResearchedProfileResult> {
const research = await getAi().models.generateContent({
model: BULK_MODEL,
contents: buildResearchPrompt(org),
config: {
tools: [{ googleSearch: {} }],
temperature: 0.2,
},
});
const researchText = research.text ?? '';
const chunks =
research.candidates?.[0]?.groundingMetadata?.groundingChunks ?? [];
const seen = new Map<string, string | null>();
for (const chunk of chunks) {
const uri = chunk.web?.uri;
if (uri != null && !seen.has(uri)) seen.set(uri, chunk.web?.title ?? null);
}
const sources = [...seen.entries()].map(([url, title]) => ({ url, title }));
if (researchText.trim() === '') {
throw new Error(
`researchOrgProfile: empty research response for "${org.name}"`,
);
}
const extraction = await getAi().models.generateContent({
model: BULK_MODEL,
contents: buildExtractionPrompt(
org,
researchText,
sources.map((s) => s.url),
),
config: {
responseMimeType: 'application/json',
responseJsonSchema: RESEARCHED_ORG_PROFILE_JSON_SCHEMA,
temperature: 0.1,
},
});
const raw: unknown = JSON.parse(extraction.text ?? '{}');
const profile = ResearchedOrgProfileSchema.parse(raw);
return { profile, sources };
}
/**
* Description-shaped text representing the researched profile in embedding
* space (RETRIEVAL_QUERY side of mission-fit). Same composition every
* re-embed so profiles are comparable run to run.
*/
export function buildProfileEmbeddingText(
org: ProfileResearchTarget,
profile: ResearchedOrgProfile,
): string {
const parts: string[] = [org.name];
if (profile.mission != null) parts.push(`Mission: ${profile.mission.value}`);
if (profile.programs.length > 0) {
parts.push(
`Programs: ${profile.programs
.map((p) =>
[p.name, p.description, p.populationServed && `serving ${p.populationServed}`]
.filter(Boolean)
.join(' — '),
)
.join('; ')}`,
);
}
if (profile.serviceGeography != null) {
parts.push(`Serves: ${profile.serviceGeography.value}`);
}
return parts.join('\n').slice(0, 8_000);
}

View File

@@ -0,0 +1,56 @@
import { describe, expect, it } from 'vitest';
import {
buildExtractionPrompt,
buildResearchPrompt,
ResearchedOrgProfileSchema,
} from './researched-profile.js';
describe('buildResearchPrompt', () => {
it('names the org, location, and identity guard', () => {
const p = buildResearchPrompt({
name: 'Seacoast Pathways',
city: 'Portsmouth',
state: 'NH',
ein: '020123456',
nteeCode: 'F80',
});
expect(p).toContain('"Seacoast Pathways"');
expect(p).toContain('Portsmouth, NH');
expect(p).toContain('020123456');
expect(p).toContain('not a similarly-named organization');
});
});
describe('buildExtractionPrompt', () => {
it('embeds sources and identity rule', () => {
const p = buildExtractionPrompt(
{ name: 'X', city: 'Keene', state: 'NH', ein: null, nteeCode: null },
'notes here',
['https://a.example', 'https://b.example'],
);
expect(p).toContain('- https://a.example');
expect(p).toContain('identityConfirmed=false');
expect(p).toContain('notes here');
});
});
describe('ResearchedOrgProfileSchema', () => {
it('accepts a minimal valid profile and rejects bad confidence', () => {
const minimal = {
identityConfirmed: false,
orgWebsite: null,
mission: null,
programs: [],
serviceGeography: null,
knownFunders: [],
staff: [],
recentNews: [],
overallConfidence: 0.1,
};
expect(ResearchedOrgProfileSchema.parse(minimal).identityConfirmed).toBe(false);
expect(() =>
ResearchedOrgProfileSchema.parse({ ...minimal, overallConfidence: 1.5 }),
).toThrow();
});
});

View File

@@ -0,0 +1,192 @@
import { z } from 'zod';
/**
* Stage 4 researched profile — aligned 1:1 with the `org_profiles` table.
* Every claim carries the URL it came from; Email 1's personalization line
* is only as trustworthy as these citations, so an uncited claim is a
* lower-trust claim by construction (docs/plan.md, Stage 4).
*/
const cited = <Value extends z.ZodTypeAny>(valueSchema: Value) =>
z.object({
value: valueSchema,
sourceUrl: z.string().nullable(),
});
export const ResearchedOrgProfileSchema = z.object({
/** Did the research confidently identify THIS org (right city/state)? */
identityConfirmed: z.boolean(),
orgWebsite: z.string().nullable(),
mission: cited(z.string().min(1)).nullable(),
programs: z.array(
z.object({
name: z.string().min(1),
description: z.string().nullable(),
populationServed: z.string().nullable(),
sourceUrl: z.string().nullable(),
}),
),
serviceGeography: cited(z.string().min(1)).nullable(),
knownFunders: z.array(
z.object({ name: z.string().min(1), sourceUrl: z.string().nullable() }),
),
staff: z.array(
z.object({
name: z.string().min(1),
role: z.string().nullable(),
sourceUrl: z.string().nullable(),
}),
),
recentNews: z.array(
z.object({
headline: z.string().min(1),
url: z.string().nullable(),
date: z.string().nullable(),
}),
),
/** Model's overall 01 confidence that the profile describes the right org accurately. */
overallConfidence: z.number().min(0).max(1),
});
export type ResearchedOrgProfile = z.infer<typeof ResearchedOrgProfileSchema>;
/** Plain JSON Schema mirror of ResearchedOrgProfileSchema for Gemini's responseJsonSchema. */
export const RESEARCHED_ORG_PROFILE_JSON_SCHEMA = {
type: 'object',
properties: {
identityConfirmed: { type: 'boolean' },
orgWebsite: { type: ['string', 'null'] },
mission: {
type: ['object', 'null'],
properties: {
value: { type: 'string' },
sourceUrl: { type: ['string', 'null'] },
},
required: ['value', 'sourceUrl'],
},
programs: {
type: 'array',
items: {
type: 'object',
properties: {
name: { type: 'string' },
description: { type: ['string', 'null'] },
populationServed: { type: ['string', 'null'] },
sourceUrl: { type: ['string', 'null'] },
},
required: ['name', 'description', 'populationServed', 'sourceUrl'],
},
},
serviceGeography: {
type: ['object', 'null'],
properties: {
value: { type: 'string' },
sourceUrl: { type: ['string', 'null'] },
},
required: ['value', 'sourceUrl'],
},
knownFunders: {
type: 'array',
items: {
type: 'object',
properties: {
name: { type: 'string' },
sourceUrl: { type: ['string', 'null'] },
},
required: ['name', 'sourceUrl'],
},
},
staff: {
type: 'array',
items: {
type: 'object',
properties: {
name: { type: 'string' },
role: { type: ['string', 'null'] },
sourceUrl: { type: ['string', 'null'] },
},
required: ['name', 'role', 'sourceUrl'],
},
},
recentNews: {
type: 'array',
items: {
type: 'object',
properties: {
headline: { type: 'string' },
url: { type: ['string', 'null'] },
date: { type: ['string', 'null'] },
},
required: ['headline', 'url', 'date'],
},
},
overallConfidence: { type: 'number', minimum: 0, maximum: 1 },
},
required: [
'identityConfirmed',
'orgWebsite',
'mission',
'programs',
'serviceGeography',
'knownFunders',
'staff',
'recentNews',
'overallConfidence',
],
} as const;
export interface ProfileResearchTarget {
readonly name: string;
readonly city: string | null;
readonly state: string;
readonly ein: string | null;
readonly nteeCode: string | null;
}
/** Prompt for the grounded research pass (Google Search tool enabled). */
export function buildResearchPrompt(org: ProfileResearchTarget): string {
const where = [org.city, org.state].filter(Boolean).join(', ');
return [
`Research the nonprofit organization "${org.name}" based in ${where}.`,
org.ein != null ? `Its federal EIN is ${org.ein}.` : '',
'',
'Find and report, with specifics:',
'1. Their mission, in their own words if possible.',
'2. Their concrete programs/services — names, what each does, and who it serves.',
'3. The geography they serve (towns/counties/statewide).',
'4. Foundations or agencies that fund them (funder walls, annual reports, press).',
'5. Key staff: executive director and anyone in a development/grants role.',
'6. Any news from the last two years.',
'',
'CRITICAL: verify you are describing THIS organization — the one in',
`${where} — not a similarly-named organization elsewhere. If you cannot`,
'find credible information clearly about this specific organization, say',
'so explicitly instead of substituting a lookalike.',
]
.filter((line) => line !== '')
.join('\n');
}
/** Prompt for the structured extraction pass over the research text. */
export function buildExtractionPrompt(
org: ProfileResearchTarget,
researchText: string,
sourceUrls: readonly string[],
): string {
return [
'Extract a structured profile of the nonprofit below from the research',
'notes. Rules:',
'- Only include claims supported by the research notes.',
'- For each claim, cite the most relevant source URL from the list, or',
' null if the claim synthesizes multiple sources.',
'- If the research failed to clearly identify this specific organization',
` (the one in ${[org.city, org.state].filter(Boolean).join(', ')}),`,
' set identityConfirmed=false and overallConfidence at or below 0.3.',
'- Empty arrays are fine. Never invent programs, funders, staff, or news.',
'',
`Organization: ${org.name}`,
'',
`Available source URLs:\n${sourceUrls.map((u) => `- ${u}`).join('\n') || '- (none)'}`,
'',
`Research notes:\n${researchText}`,
].join('\n');
}

View File

@@ -6,3 +6,5 @@ export * from './agents/org-profiler/run.js';
export * from './agents/mission-fit-judge/schema.js'; export * from './agents/mission-fit-judge/schema.js';
export * from './agents/mission-fit-judge/run.js'; export * from './agents/mission-fit-judge/run.js';
export * from './grant-embedding-text.js'; export * from './grant-embedding-text.js';
export * from './agents/org-profiler/researched-profile.js';
export * from './agents/org-profiler/research.js';

View File

@@ -24,6 +24,11 @@ export interface MatchDetail {
registrationNumber: string | null; registrationNumber: string | null;
missionStatement: string | null; missionStatement: string | null;
profileConfidence: number | null; profileConfidence: number | null;
programs: unknown;
knownFunders: unknown;
staff: unknown;
serviceGeography: string | null;
profileSources: unknown;
}; };
grant: { grant: {
id: string; id: string;
@@ -103,6 +108,11 @@ export async function serverGetMatchDetail(
.select({ .select({
missionStatement: schema.orgProfiles.missionStatement, missionStatement: schema.orgProfiles.missionStatement,
confidence: schema.orgProfiles.confidence, confidence: schema.orgProfiles.confidence,
programs: schema.orgProfiles.programs,
knownFunders: schema.orgProfiles.knownFunders,
staff: schema.orgProfiles.staff,
serviceGeography: schema.orgProfiles.serviceGeography,
sources: schema.orgProfiles.sources,
}) })
.from(schema.orgProfiles) .from(schema.orgProfiles)
.where(eq(schema.orgProfiles.orgId, row.orgId)) .where(eq(schema.orgProfiles.orgId, row.orgId))
@@ -184,6 +194,11 @@ export async function serverGetMatchDetail(
registrationNumber: row.orgRegNo, registrationNumber: row.orgRegNo,
missionStatement: profiles[0]?.missionStatement ?? null, missionStatement: profiles[0]?.missionStatement ?? null,
profileConfidence: profiles[0]?.confidence ?? null, profileConfidence: profiles[0]?.confidence ?? null,
programs: profiles[0]?.programs ?? null,
knownFunders: profiles[0]?.knownFunders ?? null,
staff: profiles[0]?.staff ?? null,
serviceGeography: profiles[0]?.serviceGeography ?? null,
profileSources: profiles[0]?.sources ?? null,
}, },
grant: { grant: {
id: row.grantId, id: row.grantId,

View File

@@ -0,0 +1,57 @@
import { sql } from 'drizzle-orm';
import type { NpOutreachDatabase, NpOutreachTransaction } from '#~/db/db.js';
import { schema } from '#~/db/db.js';
export interface ResearchedProfileInput {
readonly orgId: string;
readonly missionStatement: string | null;
readonly programs: unknown;
readonly serviceGeography: string | null;
readonly recentNews: unknown;
readonly knownFunders: unknown;
readonly staff: unknown;
readonly sources: unknown;
readonly confidence: number;
/** 1536-dim RETRIEVAL_QUERY embedding of the researched profile text. */
readonly profileEmbedding: number[];
}
/**
* Applies a Stage 4 researched profile, OVERWRITING the row — this is the
* upgrade path the v0 NTEE stub (confidence 0.2, never-overwriting) exists
* to be replaced by. Re-running research refreshes in place.
*/
export async function serverApplyResearchedProfile(
db: NpOutreachDatabase | NpOutreachTransaction,
profile: ResearchedProfileInput,
): Promise<void> {
await db
.insert(schema.orgProfiles)
.values({
orgId: profile.orgId,
missionStatement: profile.missionStatement,
programs: profile.programs,
serviceGeography: profile.serviceGeography,
recentNews: profile.recentNews,
knownFunders: profile.knownFunders,
staff: profile.staff,
sources: profile.sources,
confidence: profile.confidence,
profileEmbedding: profile.profileEmbedding,
})
.onConflictDoUpdate({
target: schema.orgProfiles.orgId,
set: {
missionStatement: sql`excluded.mission_statement`,
programs: sql`excluded.programs`,
serviceGeography: sql`excluded.service_geography`,
recentNews: sql`excluded.recent_news`,
knownFunders: sql`excluded.known_funders`,
staff: sql`excluded.staff`,
sources: sql`excluded.sources`,
confidence: sql`excluded.confidence`,
profileEmbedding: sql`excluded.profile_embedding`,
},
});
}

View File

@@ -2,3 +2,4 @@ export * from './insert-org.server.js';
export * from './upsert-org-from-registry.server.js'; export * from './upsert-org-from-registry.server.js';
export * from './enrich-org.server.js'; export * from './enrich-org.server.js';
export * from './ensure-org-profile-stub.server.js'; export * from './ensure-org-profile-stub.server.js';
export * from './apply-researched-profile.server.js';

View File

@@ -1,3 +1,4 @@
export * from './list-orgs-in-icp-band.server.js'; export * from './list-orgs-in-icp-band.server.js';
export * from './list-orgs-needing-enrichment.server.js'; export * from './list-orgs-needing-enrichment.server.js';
export * from './list-match-candidate-orgs.server.js'; export * from './list-match-candidate-orgs.server.js';
export * from './list-orgs-needing-profile.server.js';

View File

@@ -0,0 +1,55 @@
import { sql } from 'drizzle-orm';
import type { NpOutreachDatabase, NpOutreachTransaction } from '#~/db/db.js';
export interface OrgNeedingProfile {
id: string;
name: string;
city: string | null;
state: string;
ein: string | null;
nteeCode: string | null;
bestMatchScore: number;
}
/** Profiles below/at this confidence are NTEE-derived stubs, not research. */
export const STUB_CONFIDENCE_CEILING = 0.2;
/**
* The plan's cost-control rule made concrete: research only orgs that
* already have at least one non-rejected easy-win match (a plausible
* lead), best-scoring orgs first, and only when the current profile is a
* stub (or missing). Re-research after review feedback comes later.
*/
export async function serverListOrgsNeedingProfile(
db: NpOutreachDatabase | NpOutreachTransaction,
{ limit }: { limit: number },
): Promise<OrgNeedingProfile[]> {
const result = await db.execute(sql`
SELECT
o.id, o.name, o.city, o.state, o.ein, o.ntee_code,
max(m.total_score) AS best_match_score
FROM orgs o
JOIN matches m ON m.org_id = o.id
AND m.review_status != 'rejected'
AND m.easy_win = true
LEFT JOIN org_profiles p ON p.org_id = o.id
-- confidence is a float4: 0.2 stores as 0.20000000298, so a plain
-- <= 0.2 comparison (float8) excludes every stub. Epsilon absorbs it.
WHERE p.id IS NULL OR p.confidence IS NULL OR p.confidence <= ${STUB_CONFIDENCE_CEILING} + 0.001
GROUP BY o.id
ORDER BY max(m.total_score) DESC
LIMIT ${limit}
`);
const { rows } = result as unknown as { rows: Record<string, unknown>[] };
return rows.map((r) => ({
id: r.id as string,
name: r.name as string,
city: (r.city as string) ?? null,
state: r.state as string,
ein: (r.ein as string) ?? null,
nteeCode: (r.ntee_code as string) ?? null,
bestMatchScore: Number(r.best_match_score),
}));
}