Deep Searcher

Category: SDKs & Libraries
Tags: open-source, milvus, langchain, vector-search

Description

Deep Searcher is a local open-source deep research solution that integrates Milvus (an open-source vector database) and LangChain to provide advanced vector search and retrieval capabilities using open-source models. It is designed as both a Python library and a command-line tool for agentic research workflows.

Features

• Agentic RAG (Retrieval-Augmented Generation): Automates deep research by decomposing complex queries, searching across multiple document sources, and synthesizing structured reports.
• Modular Research Pipeline:
  • Define/Refine the Question: Breaks down user queries into sub-queries for more granular research.
  • Query Routing: Uses an LLM to route sub-queries to only the most relevant data sources or collections in Milvus.
  • Similarity Search: Retrieves relevant document chunks using vector search with Milvus.
  • Reflection: The agent reflects on the completeness of its answers, determining if additional sub-queries are needed.
  • Conditional Execution Flow: Automatically repeats research steps as needed, based on LLM output, until the research is complete.
  • Synthesis: Combines all findings into a consistent and well-structured report.
• Flexible Data Source Configuration: Allows input of multiple source documents and manual specification of data sources (local or online).
• Embedding Model and Vector DB Selection: Embedding model and vector database can be configured via a configuration file.
• Web Crawling as a Tool: Capable of using web crawling for additional data gathering (planned for future updates).
• Open-Source: Fully open-source and designed for local or private deployment.
• Supports Multiple Inference Services: Works with most inference services, including OpenAI, Gemini, DeepSeek, and Grok 3 (coming soon).
• Efficient Inference: Demonstrates use with fast and scalable inference services (e.g., DeepSeek-R1 on SambaNova hardware), but can also run locally with open models.

Pricing

No pricing information is specified. Deep Searcher is open-source and can be self-hosted. (Inference services such as DeepSeek-R1 may have their own costs.)

Source

Deep Searcher Blog Introduction