Token consolidation:
- docker-compose web service now reads ${CLAUDE_CODE_OAUTH_TOKEN} directly,
drop the W1-F8 CLAUDE_CODE_OAUTH_TOKEN_FOR_WEB indirection (user feedback:
one var name, no _FOR_WEB suffix).
investigator-runtime claude.ts:
- --system-prompt silently dropped by CLI v2.1.150 for multi-KB prompts;
inline the system content into the user prompt with a separator
(mirrors scripts/reextract/run.py pattern).
- Multi-line prompts via positional -- broke ("Input must be provided …");
pipe via stdin instead.
- --allowedTools "" is rejected; when no tools wanted, omit it and explicitly
--disallowedTools the writer/reader set so the model can't reach for any.
investigator-runtime locard.ts:
- Log the raw response (first 600 chars) to container stderr — saved hours
of debugging when the writer rejected.
- Grade fallback: when Locard omits `grade` but provides custody_steps,
infer the highest grade that fits (≥3 → A, ≥2 → B, ≥1 → C).
investigator-runtime write_evidence.ts:
- Filter related_hypotheses entries with empty/null hypothesis_id silently
(Locard sometimes emits [{}] when it knows no link yet) instead of
failing the whole write.
Migration 0006_investigator_serial_sequences.sql:
- BIGSERIAL on the 7 investigation tables created auto-sequences
(evidence_evidence_pk_seq etc) that 0004 forgot to GRANT to the
investigator role. Without those grants every INSERT failed with
"permission denied for sequence …". Grant USAGE/SELECT/UPDATE on each
auto-seq.
Verified live: Locard wrote E-0002 + E-0003 from real Sandia chunks
(green fireball Feb 1949; cobalt particle analysis). Grade B, confidence
high, custody chain of 3 steps with honest gaps. Cost $0.09 for both,
~70s wall.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
|
||
|---|---|---|
| .claude/agents | ||
| .forgejo/workflows | ||
| .nirvana/outputs/disclosure-bureau | ||
| case/gaps | ||
| docs/adrs | ||
| infra | ||
| investigator-runtime | ||
| scripts | ||
| tests/rag | ||
| web | ||
| wiki/entities/events | ||
| .gitignore | ||
| CHANGELOG.md | ||
| CLAUDE-schema-full.md | ||
| CLAUDE.md | ||
| CORPUS-SNAPSHOT.md | ||
| README.md | ||
The Disclosure Bureau
Investigative wiki + agentic chat sobre o corpus declassificado do US Department of War em
war.gov/ufo(116 PDFs, 3.435 páginas, 34k+ entidades, 28 vídeos UAP).
Live: disclosure.top
O que é
Pipeline de IA que transforma documentos UAP/UFO declassificados em uma wiki investigativa navegável + chat agêntico com retrieval semântico bilíngue (EN + PT-BR) e citações com bbox crop no PDF original.
A premissa metodológica é o padrão Karpathy LLM Wiki: ler tudo, compilar conhecimento em markdown cross-referenciado, navegar via wiki-links — não por busca vetorial. Em cima dessa wiki rodamos uma camada de hybrid retrieval (BM25 + BGE-M3 dense + cross-encoder rerank) para perguntas livres no chat.
A camada investigativa segue protocolo Investigation Bureau (Holmes/Poirot/Dupin/Locard + Schneier/Tetlock/Taleb): chain-of-custody, hypothesis tournament, residual uncertainty.
Arquitetura
PDFs (raw/)
↓ pdftoppm 72 DPI + pdftotext
processing/ (png + ocr)
↓ Sonnet 4.6 subagents (page-rebuilder, image-analyst, table-stitcher)
raw/<doc>--subagent/ (chunks bilíngues + bbox + anomaly flags)
↓ scripts/30 (BGE-M3 embed) + 31 (entity_mentions)
Postgres + pgvector + tsvector
↓ hybrid_search RPC + reranker
chat agente (OpenRouter) cita [[doc/p007#c0042]] → frontend renderiza crop bbox
Stack
- Embedding: BGE-M3 self-hosted (1024-dim, multilíngue, $0)
- Reranker: BGE-Reranker-v2-M3 self-hosted ($0)
- Vetor + texto: Postgres 15 + pgvector + tsvector bilíngue (
pt_unaccent,en_unaccent) - LLM (chat): OpenRouter — DeepSeek v4 free como default
- Frontend: Next.js 15 + React 19 + Tailwind + assistant-ui (Pattern C streaming)
- Auth + persistência: Supabase self-hosted (GoTrue, PostgREST, Storage, Imgproxy)
- Reverse proxy: Traefik + Let's Encrypt
- Imagens: sharp via
/api/crop(bbox on-demand, cached 1ano)
Layout
/Users/guto/ufo/
├── CLAUDE.md # contrato vinculante (24 tipos de markdown)
├── CLAUDE-schema-full.md # schema detalhado
├── README.md # este arquivo
├── raw/ # 116 PDFs imutáveis + chunks v0.2.0 derivados
│ ├── <pdfs>
│ ├── <doc-id>--subagent/ # chunks rebuilt (chunks/c*.md + _index.json + document.md)
│ └── _batch-rebuild/ # logs do orchestrator
├── processing/ # intermediários (PNG, OCR, vision JSON)
├── wiki/ # markdown gerado (documents/, pages/, entities/, tables/, images/)
├── case/ # artefatos Investigation Bureau (case-report, hypotheses, gaps)
├── scripts/ # 33 scripts numerados (Phase 0 → manutenção)
├── infra/ # docker-compose, embed-service, migrations, deploy
└── web/ # Next.js frontend
Quick start
# 1. Converter PDFs em PNG + OCR (uma vez)
./scripts/01-convert-pdfs.sh
# 2. Rebuild chunks bilíngues (Sonnet 4.6 via Claude Code subagents)
python3 scripts/28-batch-rebuild-all.py --workers 2
# 3. (após batch) Indexar em Postgres + embeddings
python3 scripts/30-index-chunks-to-db.py --skip-existing
# 4. (opcional) Materializar entity_mentions p/ grafo
python3 scripts/31-populate-entity-mentions.py
# 5. Deploy
cd infra/disclosure-stack && ./scripts/deploy.sh
Detalhes completos em infra/DEPLOY-CHECKLIST.md.
Features do frontend
| URL | Função |
|---|---|
/ |
Lista de documentos com resumo de 3 linhas, filtros (collection, classification, sort), busca |
/d/<doc> |
Visão legado (page grid + frontmatter) |
/d/<doc>/v2 |
Render rico de chunks com lang toggle (PT/EN/both), paged vs flow |
/d/<doc>/v2/<page> |
Single page V2 com PNG side-by-side |
/d/<doc>/full |
Texto consolidado bilíngue |
/e/<class> |
Lista paginada de entidades por classe (people, locations, ...) |
/e/<class>/<id> |
Detalhe da entidade + co-mentions + chunks live |
/search?q=... |
Hybrid search URL-shareable |
/timeline |
Cronologia de eventos por década |
/graph |
Grafo força-direcionado de co-menções (Obsidian-style) |
/admin/stats |
Analytics do corpus (FS + DB) |
/admin/batch |
Monitor de progresso do rebuild |
/admin/indexer |
Estado da camada de retrieval |
Atalhos globais:
⌘K/Ctrl+Kem qualquer página → command palette com hybrid_search- Toggle 🌐 EN ↔ PT-BR fixo bottom-left (cookie 1ano)
- Chat 💬 botão flutuante bottom-right com 12 ferramentas
Os 12 tools do agente
🔍 Retrieval: hybrid_search, read_chunk, get_page_chunks, list_anomalies
🔗 Grafo: entity_neighbors, entity_path, co_mention_chunks
📄 Wiki: read_document, read_page, read_entity, search_corpus
🧭 UI: navigate_to
Citações tipo [[doc-id/p007#c0042]] viram cards interativos com crop bbox + texto bilíngue + link.
Custos
| Item | Custo |
|---|---|
| Rebuild chunks (Sonnet 4.6 via Claude Code Max 20x) | ~$200 one-shot p/ 116 docs |
| Embedding BGE-M3 self-host | $0/mês |
| Reranker BGE-Reranker-v2-M3 self-host | $0/mês |
| Postgres + pgvector | já incluso no VPS |
| Chat LLM (DeepSeek free via OpenRouter) | $0/req |
| VPS (16GB / 4 CPU) | ~€10/mês |
Documentação
CLAUDE.md— contrato vinculante para agentes (schema v0.2.0)CLAUDE-schema-full.md— schema dos 24 tiposinfra/RETRIEVAL.md— arquitetura da camada de retrievalinfra/DEPLOY-CHECKLIST.md— runbook end-to-endinfra/embed-service/README.md— microsserviço BGE-M3
Licença + procedência
- PDFs declassificados: domínio público (US Department of War / FBI / DOS / NASA)
- Código deste projeto: MIT
- Modelos: BGE-M3 (MIT), DeepSeek v4 (proprietary via OpenRouter free tier)
- Branding: The Disclosure Bureau / disclosure.top — pessoal
Wiki investigativa, não advocacy. Toda claim tem chain-of-custody até a página + bbox do PDF original.