Forager – Colheitor de Conhecimento Slack
dallascrilley/forager-demo
Transforma conhecimento tribal do Slack em Q&A com confiança scored
Ver no GitHubMCP server que converte threads resolvidas do Slack em uma base de Q&A consultável com pontuação de confiança. Ingere exportações reais de canais Slack, agrupa mensagens em threads e extrai pares pergunta-resposta pontuados por sinais sociais (reações, confirmações). Prova híbrida com backend funcional e ingestion server-side real, sem OAuth ou busca semântica.
Instalação
npx convez-ai dallascrilley/forager-demo1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Extrair FAQs e respostas resolvidas de histórico de Slack corporativo
- Criar base de conhecimento consultável a partir de threads com reações/confirmações
- Documentar decisões e soluções técnicas capturadas em canais de resolução
Veredito

Prova bem executada de pipeline de extração de conhecimento com ingestion real e scoring honesto baseado em evidência social. Útil para equipes que querem mineração de Slack sem OAuth, mas com fronteiras claras sobre o que está fora de escopo (embeddings, conexão live).
README
A tradução para o português deste README ainda está na fila.
Forager
Your tribal knowledge is in Slack. Stop pretending the wiki has it.
Forager turns resolved Slack threads into a queryable Q&A knowledge base with confidence scoring. It is a hybrid proof: a real backend ingests an actual Slack export you upload — grouping messages into threads, harvesting question/answer pairs, and scoring each by social signal — while live OAuth and semantic search are explicitly out of scope.
Live demo: dallascrilley.com/demos/forager — explore the synthetic sample, or upload a real Slack channel export and watch it get harvested server-side.
Real vs. synthetic — the honest boundary
| Capability | Source |
|---|---|
Ingest a real Slack channel export (ts/thread_ts/reactions) | Live — server-side parsing of your uploaded JSON |
| Thread grouping, Q&A harvesting, confidence scoring | Live — runs in the backend on real input |
| Sample workspace (no upload) | Synthetic — public/data/threads.json |
| Live Slack connection | Out of scope — no OAuth / Event API |
| Semantic retrieval | Out of scope — matching is keyword-based, not embeddings |
The synthetic sample lets a reviewer try it instantly; uploading a real export proves the ingestion and harvesting logic works on genuine workspace data.
The backend
functions/forager/ingest.js is a Cloudflare Pages Function — POST /forager/ingest. It:
- accepts either a raw Slack channel export (
[{ ts, text, thread_ts?, reactions?, user_profile? }]) or the demo's normalized thread shape, and auto-detects which; - groups messages into threads by
thread_ts, then harvests Q&A: detect questions, resolve each to its confirmed answer (reaction/thanks signal, or longest substantive reply), and score confidence from social signal — resolution status, 👍/🔥 reactions, confirmation language — capped at 0.98 because extracted knowledge is never certain; - is stateless — the request body is the only input; nothing is stored.
curl -X POST https://dallascrilley.com/demos/forager/ingest \
-H 'content-type: application/json' \
-d '{"raw":"[{\"ts\":\"1700000000.0001\",\"text\":\"How do we rotate DKIM keys?\"}]"}'
The ingestion and scoring logic are pure functions, exported and unit-tested in tests/ingest.test.js.
Run locally
pnpm install
pnpm test # unit tests for harvesting + scoring
pnpm dev # static UI only — http://localhost:4321 (synthetic sample)
pnpm build && npx wrangler pages dev dist # UI + live /forager/ingest — http://localhost:8788
Uploads reach the backend only under wrangler pages dev (port 8788); pnpm dev (port 4321) serves the synthetic UI alone.
What it proves
- Knowledge-extraction pipeline design — question detection → answer resolution → confidence scoring, on real Slack-export structure.
- Defensive ingestion — two input shapes auto-detected, malformed input rejected with clear errors.
- Confidence from evidence, not vibes — scores derive from reactions and resolution signals and never reach certainty.
- Honest system boundaries — the live/synthetic and "no OAuth, no embeddings" lines are explicit in the UI, the API response (
source,inputFormat), and this README.
Architecture
See ARCHITECTURE.md for the data model, harvester pipeline, scoring formula, backend design, and tradeoffs.
License
MIT