Roadmap¶
v0.1 — foundations ✅¶
- Loaders: PDF, Markdown, TXT, HTML, DOCX; page-aware chunking; idempotent ingest
- Vector search (pgvector HNSW) with cited passages
- CLI, REST API, MCP (stdio + Streamable HTTP), Docker Compose
v0.2 — quality ✅¶
- Hybrid search: pgvector + PostgreSQL full-text (Turkish stemming), RRF fusion
- Optional cross-encoder rerank (
bge-reranker-v2-m3) - Eval harness (Recall@k, MRR, page hit, latency) and a Turkish sample set in CI
v0.3 — MCP & multi-tenant¶
- [x] Multiple collections with per-collection language
- [x] MCP resources (
citewell://…) and theanswer_with_citationsprompt - [ ] Argument completions for collection names
- [ ] Progress notifications for long operations
- [ ] Per-collection tokens / read scopes for the HTTP endpoint
Later¶
- OpenAI-compatible
/v1/chat/completionsadapter so UIs like Open WebUI can use Citewell (generation delegated to a user-configured model; core stays LLM-free) - OCR for scanned PDFs (Tesseract
tur) - Accent-insensitive Turkish matching; sparse / BM25 retriever as a third ranking
- Public Turkish retrieval benchmark
- SQLite single-file mode for zero-dependency installs