local-deep-researcher

Local Deep Researcher is an open-source, fully local web research and report-writing assistant built on LangGraph. It runs entirely on your machine using any LLM hosted by Ollama or LMStudio, requiring no cloud-based AI services. **How It Works** The assistant follows an iterative research loop inspired by the IterDRAG paper: 1. **Query Generation**: Given a user-provided topic, a local LLM generates a web search query. 2. **Web Search**: A configurable search engine (DuckDuckGo by default, or SearXNG, Tavily, or Perplexity) retrieves relevant sources. 3. **Summarization**: The LLM summarizes the search results in relation to the research topic. 4. **Gap Analysis**: The LLM reflects on the summary to identify knowledge gaps. 5. **New Query**: A new search query is generated to address the identified gaps. 6. **Iteration**: Steps 2–5 repeat for a user-defined number of cycles (default: 3). The final output is a markdown file containing the research summary with citations to all sources used. **Model Support** - **Ollama**: Supports any model available on Ollama's library (e.g., DeepSeek R1, Llama 3.2). Configuration via `OLLAMA_BASE_URL` and `LOCAL_LLM` environment variables. - **LMStudio**: Supports models loaded in LMStudio with an OpenAI-compatible API server. Configuration via `LMSTUDIO_BASE_URL` and `LOCAL_LLM`. - **Tool Calling**: As of August 2025, supports tool calling for models like gpt-oss that do not support JSON mode in Ollama. **Search Engine Options** - **DuckDuckGo** (default): No API key required. - **SearXNG**: Self-hosted metasearch engine. - **Tavily**: Requires API key. - **Perplexity**: Requires API key. **Deployment** - **LangGraph Studio**: Launch locally with `langgraph dev` for an interactive UI where you can configure settings and visualize the research process step by step. - **Docker**: A Dockerfile is provided, but Ollama must be run separately as a dependent service. - **Configuration Priority**: Environment variables > LangGraph UI configuration > default values in the Configuration class. **Technical Details** - Built with LangGraph for stateful, multi-step agent workflows. - Uses `python-dotenv` for environment variable management. - Includes fallback mechanisms for models that struggle with structured JSON output (e.g., DeepSeek R1 7B and 1.5B). - A TypeScript port is available at a separate repository. - Browser compatibility: Firefox recommended; Safari may encounter mixed-content warnings. **Use Cases** - Deep research on any topic without sending data to cloud AI services. - Generating cited research reports for personal or professional use. - Exploring topics iteratively with gap-aware search strategies. - Learning how agentic AI research workflows function.