Bug Description

The @pleaseai/context-please-mcp v0.5.0 package ignores the VECTOR_DB_TYPE=milvus environment variable and continues to use FAISS for vector storage, despite having Milvus correctly configured and running.

Environment

• Package: @pleaseai/context-please-mcp@0.5.0
• Node.js: v20.17
• Milvus: v2.4.1 (self-hosted, running in Docker)
• Embedding Provider: OpenAI (using SambaNova API endpoint)
• Platform: Linux (DDEV development environment)

Configuration

{
  "context-please": {
    "type": "stdio",
    "command": "npx",
    "args": ["-y", "@pleaseai/context-please-mcp@latest"],
    "env": {
      "VECTOR_DB_TYPE": "milvus",
      "FAISS_STORAGE_DIR": "/home/lucas/.context/faiss-indexes/",
      "EMBEDDING_PROVIDER": "OpenAI",
      "OPENAI_BASE_URL": "https://api.sambanova.ai/v1",
      "OPENAI_API_KEY": "${SAMBANOVA_API_KEY}",
      "EMBEDDING_MODEL": "E5-Mistral-7B-Instruct",
      "NODE_ENV": "production",
      "MILVUS_ADDRESS": "milvus:19530"
    }
  }
}

Steps to Reproduce

1. Set up Milvus:

   # Milvus 2.4.1 running in Docker
   # Health check: http://milvus:9091/healthz returns "OK"
   # Accessible at: milvus:19530

2. Configure MCP with Milvus:

   • Set VECTOR_DB_TYPE=milvus
   • Set MILVUS_ADDRESS=milvus:19530
   • Start the MCP server

3. Verify environment variables (checking running process):

   cat /proc/<pid>/environ | tr '\0' '\n' | grep -E "(VECTOR|MILVUS)"

   Output:

   VECTOR_DB_TYPE=milvus
   MILVUS_ADDRESS=milvus:19530
   

   ✅ Environment variables are correctly set

4. Index a codebase:

   mcp-cli call context-please/index_codebase '{"path": "/path/to/code", "splitter": "ast"}'

5. Check where data is stored:

   # Check FAISS directory
   ls -lh ~/.context/faiss-indexes/
   # Output: Contains index files (dense.index, sparse.json, documents.json)
   
   # Check Milvus collections
   python3 <<PYTHON
   from pymilvus import connections, utility, Collection
   connections.connect(host='milvus', port='19530')
   for coll in utility.list_collections():
       c = Collection(coll)
       print(f"{coll}: {c.num_entities} entities")
   PYTHON
   # Output: Collections exist but have 0 entities

Expected Behavior

When VECTOR_DB_TYPE=milvus is set:

• Data should be stored in Milvus collections
• FAISS should not be used
• Milvus collections should contain indexed data
• Search should query Milvus

Actual Behavior

Despite VECTOR_DB_TYPE=milvus:

• ❌ Data is stored in ~/.context/faiss-indexes/
• ❌ Milvus collections are created but remain empty (0 entities)
• ❌ Search fails with: "Collection exists but statistics could not be retrieved from the vector database"
• ✅ Indexing reports success: "541 files, 2637 chunks"
• ✅ Environment variables are correctly set in the process

Evidence

1. Process Environment Variables

$ cat /proc/28551/environ | tr '\0' '\n' | grep -E "(VECTOR|MILVUS)"
EMBEDDING_MODEL=E5-Mistral-7B-Instruct
EMBEDDING_PROVIDER=OpenAI
FAISS_STORAGE_DIR=/home/lucas/.context/faiss-indexes/
MILVUS_ADDRESS=milvus:19530
VECTOR_DB_TYPE=milvus  ← CORRECTLY SET

2. Data Location

$ ls -lh ~/.context/faiss-indexes/hybrid_code_chunks_*/
total 1.2M
-rw-r--r-- 1 lucas lucas 1.1M dense.index      ← Data in FAISS
-rw-r--r-- 1 lucas lucas 121K documents.json
-rw-r--r-- 1 lucas lucas  151 metadata.json
-rw-r--r-- 1 lucas lucas  55K sparse.json

3. Milvus Collections

# Collections created but empty
hybrid_code_chunks_40d1b2d8:
  Entities: 0              ← NO DATA
  Indexes:
    - vector: AUTOINDEX
    - sparse_vector: SPARSE_INVERTED_INDEX

4. Search Error

Error: Collection exists for '/var/www/html' but statistics could not be retrieved from the vector database.

Root Cause Analysis

The MCP server appears to:

1. Read VECTOR_DB_TYPE=milvus configuration ✅
2. Create Milvus collections (with correct schema) ✅
3. But then use FAISS for actual data storage ❌
4. Try to read statistics from Milvus (which is empty) during search ❌

This suggests a conditional logic bug where the vector DB type check isn't being applied consistently across:

• Collection creation (uses Milvus)
• Data insertion (uses FAISS)
• Data retrieval (tries Milvus)

Additional Notes

Sparse Vector Index Issue (Solved)

When Milvus is properly used, there's a secondary issue: self-hosted Milvus 2.4.x requires manual creation of sparse vector indexes. I've created a workaround:

from pymilvus import connections, Collection, utility

connections.connect(host='milvus', port='19530')
for coll_name in utility.list_collections():
    coll = Collection(coll_name)
    indexes = {idx.field_name for idx in coll.indexes}

    if 'sparse_vector' not in indexes:
        coll.create_index(
            field_name="sparse_vector",
            index_params={
                "metric_type": "IP",
                "index_type": "SPARSE_INVERTED_INDEX",
                "params": {"inverted_index_algo": "DAAT_WAND"}
            },
            index_name="sparse_inverted_index"
        )
        coll.load()

This is a separate issue but worth noting for self-hosted Milvus deployments.

Workaround

For now, using VECTOR_DB_TYPE=faiss-local works perfectly:

• Indexing succeeds
• Search works
• No Milvus dependency issues

Impact

• Severity: High - Core feature (Milvus integration) doesn't work
• Workaround: Use FAISS (works fine for smaller codebases)
• User Impact: Users cannot use production-grade Milvus for large codebases

System Information

$ npm view @pleaseai/context-please-mcp version
0.5.0

$ node --version
v20.17.0

$ python3 -c "from pymilvus import __version__; print(__version__)"
2.6.5

Related

• Milvus Documentation: https://milvus.io/docs/v2.4.x/sparse_vector.md
• MCP + Milvus: https://milvus.io/docs/milvus_and_mcp.md

Thank you for this excellent tool! Happy to provide more information or test fixes.