Item 7: Vector database + RAG cartridges wave (epic #87)

[hyperpolymath]

Tracker for epic #87 item 7 — multi-cartridge wave.

Scope

Four cartridges, all in the same domain (vector storage + RAG memory):

• `pinecone-mcp` — Pinecone hosted vector DB (managed, serverless)
• `weaviate-mcp` — Weaviate (open-source, self-hostable or cloud)
• `qdrant-mcp` — Qdrant (Rust-native, self-host or cloud)
• `chromadb-mcp` — Chroma (embedded or client/server; LLM-app focus)

Why one issue covers four

The four share a domain shape (vector index → upsert → query → delete + collection management) and likely share most of the manifest/tool surface. Worth designing the operation taxonomy once across all four, then implementing per-provider. The cartridge architecture lets each ship independently once the shared manifest pattern is set.

Surface

Per cartridge, ~7 operations: `authenticate`, `list_collections`, `create_collection`, `upsert`, `query` (similarity search), `delete`, `delete_collection`.

A single bridge tool `boj_vector` (mirroring `boj_search`) dispatches via `{operation, provider, collection, vectors|query|ids, params}`. Same shape as the just-shipped `search-mcp` in PR #99.

Composition with existing cartridges

• Pairs with `local-memory-mcp` (already in repo, claims 13 tools) for hybrid local-cloud memory
• Pairs with `codeseeker-mcp` for code-specific RAG → covered, complementary
• Pairs with the new `boj_search` (PR feat: item 8 (search-mcp cartridge) + ADR-0013 streamable HTTP transport (epic #87) #99) for retrieval-augmented Q&A workflows
• Composes inside prompt templates from PR feat(mcp-bridge): implement resources + prompts surface (epic #87 item 1) #89 — a future `rag-query` prompt template orchestrates search → upsert → query → answer

Implementation plan

Each provider as a sibling PR:

1. `feat/item-7a-pinecone-mcp` — Pinecone, simplest API; landed first as the shape-setting reference
2. `feat/item-7b-qdrant-mcp` — Qdrant, popular self-host option
3. `feat/item-7c-weaviate-mcp` — Weaviate, hybrid (vector + keyword) search story
4. `feat/item-7d-chromadb-mcp` — Chroma, embedded option, last because it adds embedded-storage complexity

~1 day per cartridge for the bridge surface + manifest + offline-menu wiring (mirroring the search-mcp pattern). Backend dispatch deferred per the cartridge-architecture norm.

Dependencies / non-dependencies

• Not blocked on item 14 (HTTP transport) — these are pure HTTP-API cartridges, work on stdio fine
• Pairs cleanly with item 9 (multi-modal) for vision-RAG (image embeddings)
• Independent of items 11 / 12 (formal verification campaigns)

Out of scope

• Local-only vector stores (FAISS, hnswlib) — these are libraries, not gateways; bare `ml-mcp` could expose them if needed but it's a different cartridge shape
• Embedding generation — that lives in `ml-mcp` (Hugging Face) or a dedicated embeddings cartridge, not in vector-DB cartridges

Exit criteria

Close when all 4 cartridges have manifest + bridge surface + offline-menu entry, mirroring the search-mcp pattern. Backend implementation tracked separately per cartridge.