fr0stykiller/dnd-chatbot
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

feat: migrate ChromaDB to HTTP server mode #1

Merged
fr0stykiller merged 2 commits from feat/chroma-server into main 2026-03-13 23:52:34 +01:00
Conversation 0 Commits 2 Files changed 9 +486 -344
fr0stykiller commented 2026-03-13 22:02:28 +01:00
Owner
Copy link

Summary

  • Replace local persistent directory ChromaDB with a dedicated chromadb/chroma container exposing HTTP API
  • Switch ingest.py and chat.py to use chromadb.HttpClient instead of local directory
  • Expose ChromaDB on host port 8100 for external access
  • Simplify entrypoint.sh and config.py

Test plan

  • ChromaDB server starts and accepts connections
  • Ingest service indexes 77 chunks from 16 files
  • Chat service connects and serves Gradio UI
  • Verify chat response quality
## Summary - Replace local persistent directory ChromaDB with a dedicated chromadb/chroma container exposing HTTP API - Switch ingest.py and chat.py to use chromadb.HttpClient instead of local directory - Expose ChromaDB on host port 8100 for external access - Simplify entrypoint.sh and config.py ## Test plan - [x] ChromaDB server starts and accepts connections - [x] Ingest service indexes 77 chunks from 16 files - [x] Chat service connects and serves Gradio UI - [x] Verify chat response quality
Replace local persistent directory ChromaDB with a dedicated chromadb/chroma
container exposing HTTP API. This enables external services (like Nanobot)
to query the vector store over the network.

Changes:
- Add chromadb service to docker-compose
- Switch ingest.py and chat.py to use chromadb.HttpClient
- Update config.py: replace CHROMA_DIR with CHROMA_URL
- Simplify entrypoint.sh with ChromaDB health wait
- Remove chroma_data volume mount from ingest/chat containers
- Expose ChromaDB on host port 8100

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Three RAG quality improvements:

1. Hybrid search (BM25 + Vector + RRF fusion)
   - BM25 catches exact keyword matches (proper nouns, names)
   - Vector search catches semantic similarity
   - RRF merges both rankings

2. Contextual chunk prefixes (Anthropic technique)
   - LLM generates context prefix per chunk at ingest time
   - Prefix describes: source document, characters, locations, events
   - Reduces failed retrievals by ~49%

3. Cross-encoder reranking (bge-reranker-v2-m3)
   - Retrieves top-20 candidates, reranks to top-k
   - Cross-encoder scores (query, doc) pairs together
   - Much more accurate than cosine similarity alone

New files: retriever.py, contextual.py
New deps: rank-bm25, sentence-transformers

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
fr0stykiller force-pushed feat/chroma-server from 8570b3384e to 30bc7b927b 2026-03-13 23:16:37 +01:00 Compare
fr0stykiller deleted branch feat/chroma-server 2026-03-13 23:52:34 +01:00
Sign in to join this conversation.
Reviewers
No reviewers
Labels
Clear labels
No items
No labels
Milestone
Clear milestone
No items
No milestone
Projects
Clear projects
No items
No project
Assignees
Clear assignees
No assignees
1 participant
Notifications
Due date
The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set.

Reference
fr0stykiller/dnd-chatbot!1
No description provided.