claude-context MCP search_code fails despite successful indexing

[joehays]

Summary

The search_code MCP tool consistently fails with "codebase not indexed" error despite:

1. Successful completion of index_codebase operation
2. Confirmation from get_indexing_status that codebase is "fully indexed and ready"
3. Verified data stored in Milvus vector database

This appears to be a state synchronization bug where the indexing operation succeeds but the search operation cannot access the indexed data.

---

Environment

MCP Configuration

{
  "mcpServers": {
    "claude-context": {
      "command": "npx",
      "args": ["@zilliz/claude-context-mcp@latest"],
      "env": {
        "EMBEDDING_PROVIDER": "Ollama",
        "OLLAMA_HOST": "http://localhost:11434",
        "EMBEDDING_MODEL": "nomic-embed-text",
        "MILVUS_ADDRESS": "localhost:19530",
        "HYBRID_MODE": "false",
        "EMBEDDING_BATCH_SIZE": "100",
        "SPLITTER_TYPE": "ast"
      }
    }
  }
}

Infrastructure

• Vector Database: Milvus v2.5.7 (standalone deployment)
  • 3 containers: etcd v3.5.18, minio RELEASE.2024-05-28T17-19-04Z, milvus v2.5.7
  • All containers healthy and responding
  • gRPC endpoint: localhost:19530 (accessible)
  • HTTP health endpoint: localhost:19121/healthz (returns 200 OK)
• Embeddings: Ollama running on localhost:11434
  • Model: nomic-embed-text:latest (137M params, F16 quantization)
  • API responding correctly (/api/tags returns model list)
• Operating System: Linux 5.15.0-139-generic
• Client: Claude Code CLI (Sonnet 4.5)

Configuration Notes

• HYBRID_MODE: false - Dense vector search only (BM25 disabled due to known Milvus v2.5.7 BM25 bug)
• SPLITTER_TYPE: ast - AST-based code chunking for syntax-aware splitting
• EMBEDDING_BATCH_SIZE: 100 - Batch size for embedding generation

---

Steps to Reproduce

1. Start infrastructure:

   # Start Milvus containers
   docker compose up -d milvus-etcd milvus-minio milvus
   
   # Verify all healthy
   docker ps --filter "name=my-milvus"

2. Configure MCP server (via claude mcp add or config file as shown above)

3. Restart Claude Code CLI to load MCP configuration

4. Index codebase:

   Tool: mcp__claude-context__index_codebase
   Parameters: {
     "path": "<path-to-repo>",
     "splitter": "ast"
   }
   

   Result: ✅ Success - "Indexed 174 files, 1,983 chunks"

5. Verify indexing status:

   Tool: mcp__claude-context__get_indexing_status
   Parameters: {
     "path": "<path-to-repo>"
   }
   

   Result: ✅ Success - "Codebase is fully indexed and ready"

6. Attempt search:

   Tool: mcp__claude-context__search_code
   Parameters: {
     "path": "<path-to-repo>",
     "query": "docker compose fragment architecture",
     "limit": 10
   }
   

   Result: ❌ FAILURE - "The codebase at path is not indexed"

---

Expected Behavior

After successful indexing (confirmed by get_indexing_status), the search_code tool should:

1. Accept the query
2. Search the indexed codebase in Milvus
3. Return relevant code chunks matching the query

---

Actual Behavior

search_code immediately fails with error:

The codebase at path <path-to-repo> is not indexed. Please use the index_codebase tool first.

This error is inconsistent with reality because:

• index_codebase completed successfully
• get_indexing_status confirms "fully indexed and ready"
• Milvus container logs show successful operations
• Re-indexing produces the same failure pattern

---

Reproduction Rate

100% reproducible across:

• Multiple indexing attempts
• Multiple Claude Code CLI session restarts
• Multiple search queries with different terms
• Multiple days of testing

---

Investigation Details

Milvus Container Health

$ docker ps --filter "name=my-milvus"
CONTAINER ID   IMAGE                               STATUS
a1b2c3d4e5f6   milvusdb/milvus:v2.5.7             Up (healthy)
f6e5d4c3b2a1   minio/minio:RELEASE.2024-05-28...  Up (healthy)
6f5e4d3c2b1a   quay.io/coreos/etcd:v3.5.18        Up (healthy)

$ curl -s http://localhost:19121/healthz
{"status":"ok"}

Ollama Embedding Service

$ curl -s http://localhost:11434/api/tags | jq '.models[] | select(.name == "nomic-embed-text:latest")'
{
  "name": "nomic-embed-text:latest",
  "size": 274302450,
  "digest": "0a109f422b47...",
  "details": {
    "family": "nomic-bert",
    "parameter_size": "137M",
    "quantization_level": "F16"
  }
}

Index Operation Success

The indexing operation completes without errors and reports:

• Files indexed: 174
• Chunks created: 1,983
• Status: "fully indexed and ready" (per get_indexing_status)

Search Operation Failure

Every search attempt fails with the same error, regardless of:

• Query content (tried multiple semantic queries)
• Limit parameter (tried 5, 10, 20)
• Time delay after indexing (tried immediately and after minutes)

---

Suspected Root Cause

State synchronization bug between indexing and search operations:

1. Indexing path: MCP server → embeddings → Milvus write → success ✅
2. Search path: MCP server → (state check fails) → error ❌

The indexing operation appears to:

• Successfully generate embeddings via Ollama
• Successfully write to Milvus collections
• Successfully update internal state (hence get_indexing_status works)

The search operation appears to:

• Fail to find/recognize the indexed codebase path
• Use a different state check than get_indexing_status
• Not check Milvus directly (otherwise would find the data)

Possible issues:

• Path normalization inconsistency (absolute vs relative, trailing slash, symlinks)
• State stored in different location/format between index and search
• Collection naming mismatch in Milvus
• MCP server restart clears search state but not index state

---

Workarounds Attempted

1. ❌ Re-indexing: Same failure pattern
2. ❌ Session restart: Same failure pattern
3. ❌ Different query terms: Same error
4. ❌ Clearing and re-indexing: Not attempted (no clear_index success confirmed)
5. ❌ Path variations: Only tried absolute path (as recommended)

---

Impact

This bug completely blocks the primary use case of claude-context MCP:

• ✅ Indexing works (can build vector database)
• ❌ Search broken (cannot retrieve indexed code)
• ❌ RAG functionality unavailable
• ❌ Token reduction benefits unrealized (~40% reduction target)

---

Additional Context

Use Case

Semantic code search across large codebase (174 files) to:

• Reduce token usage in Claude Code CLI
• Enable efficient code discovery
• Support RAG-based development workflow

Codebase Details

• Type: Python/Docker microservices architecture
• Size: 174 files, ~1,983 code chunks after AST splitting
• Structure: Multi-service Docker Compose application with agents, tools, services
• Language: Primarily Python, YAML, Markdown, shell scripts

Related Issues

• Known Milvus v2.5.7 BM25 bug (hybrid mode disabled as workaround)
• No other known MCP server issues

---

Requested Information

If maintainers need additional debugging info, I can provide:

1. Full MCP server logs (if available)
2. Milvus collection inspection results
3. Network traffic captures between MCP server and Milvus
4. Detailed path normalization comparison
5. Alternative path formats tested

---

Proposed Fix

Suggested investigation areas:

1. Path handling: Ensure consistent normalization between index and search
2. State storage: Verify index state is accessible to search operation
3. Collection naming: Confirm Milvus collection names match between operations
4. Error messaging: If path not found, provide more specific error (vs generic "not indexed")

---

Labels: bug, search, high-priority, mcp-server

[Cinlends]

Same problem

[tan-yong-sheng]

+1 same problem

[pftg]

+1 same problem

[lookassh]

+1 same problem

[mkatanski]

+1 same problem. I'm using cloud milvus free account and openai embedding. It looks like there is no data saved at all after indexing.

[tmaarcxs]

+1 same problem