mcp-server-ollama-deep-researcher

# Ollama Deep Researcher Architecture ## Overview The Ollama Deep Researcher is a Model Context Protocol (MCP) server that provides deep research capabilities using local LLMs via Ollama. The system integrates multiple components to enable efficient research, tracing, and monitoring. ## Core Components ### 1. MCP Resources - Persistent storage of research results - Resource URI system for accessing past research - Structured content display in MCP clients - Efficient context management through resource references ### 2. Research Engine - Implements iterative research workflow using LangGraph - Manages state transitions between research phases - Handles query generation and result synthesis ### 2. Search Integration - Supports multiple search providers (Perplexity, Tavily) - Abstracts search API interactions - Handles rate limiting and error recovery ### 3. LLM Integration - Connects to Ollama for local LLM access - Manages model loading and inference - Handles prompt engineering and response parsing ### 4. Tracing System - LangSmith integration for comprehensive tracing - Monitors all LLM operations and search requests - Provides performance metrics and debugging tools ## Data Flow ```mermaid graph TD A[User Query] --> B[Query Generator] B --> C[Search Engine] C --> D[Result Processor] D --> E[Summary Generator] E --> F[Knowledge Gap Analyzer] F --> G{Continue Research?} G -->|Yes| B G -->|No| H[Final Summary] subgraph "Tracing Layer" I[LangSmith Tracer] I -.-> B I -.-> C I -.-> D I -.-> E I -.-> F end ``` ## Key Features ### 1. Configurable Research Parameters - Maximum research loops - LLM model selection - Search API selection - Timeout settings ### 2. Performance Monitoring - Operation latency tracking - Resource utilization metrics - Error rate monitoring - Query performance analysis ### 3. Error Handling - Graceful degradation - Automatic retries - Fallback strategies - Clear error reporting ## Implementation Details ### Resource Management 1. Storage - Research results stored with URI-safe topic names - JSON-formatted content with metadata - Timestamp tracking for result freshness 2. Access - Dynamic URI templates for flexible retrieval - Resource listing for discovery - Error handling for missing resources 3. Integration - Automatic resource creation after research - Resource panel display in MCP clients - Efficient context reuse in conversations ### Research Workflow 1. Query Generation - Uses LLM to create targeted search queries - Incorporates context from previous iterations - Tracked via LangSmith for optimization 2. Web Search - Configurable search provider - Result deduplication - Source validation - Performance monitoring 3. Result Synthesis - Incremental summary updates - Source tracking - Citation management - Quality metrics ### Tracing Implementation The system uses LangSmith for comprehensive tracing: 1. Operation Tracing ```python @traceable def tavily_search(query, include_raw_content=True, max_results=3): tavily_client = TavilyClient() return tavily_client.search(query, max_results=max_results, include_raw_content=include_raw_content) ``` 2. Performance Monitoring - Response time tracking - Resource utilization - Error rate monitoring - Query optimization 3. Debug Capabilities - Step-by-step workflow tracing - Input/output inspection - Error context capture - Performance bottleneck identification ## Configuration ### Environment Variables ```bash LANGSMITH_TRACING=true LANGSMITH_ENDPOINT="https://api.smith.langchain.com" LANGSMITH_API_KEY="your-api-key" LANGSMITH_PROJECT="ollama-deep-researcher-mcp-server" TAVILY_API_KEY="tvly-your-key" # Include tvly- prefix PERPLEXITY_API_KEY="pplx-your-key" ``` ### Default Settings ```json { "maxLoops": 3, "llmModel": "deepseek-r1:1.5b", "searchApi": "perplexity", "timeout": 300 } ``` ## Future Enhancements 1. Advanced Tracing - Custom trace aggregation - Advanced performance analytics - Automated optimization suggestions 2. Search Optimization - Dynamic provider selection - Improved result ranking - Enhanced deduplication 3. LLM Improvements - Model performance tracking - Automated prompt optimization - Response quality metrics 4. System Scalability - Parallel research workflows - Distributed tracing - Enhanced resource management