DeepSearcher

DeepSearcher is an open-source, Apache-2.0 licensed Python project from Zilliz. It combines large language models with vector databases such as Milvus and Zilliz Cloud to search, evaluate and reason over private data, returning answers and comprehensive reports. The project describes itself as an open-source deep research alternative.

What It Is

DeepSearcher is a retrieval-augmented deep research library and service. Users load local files (PDF, text and markdown) or crawled web content into a vector database, then ask a question or request a report. The project positions it for enterprise knowledge management, intelligent Q&A systems and information retrieval.

How the Workflow Runs

You install it with pip, configure an LLM, an embedding model and a vector database through a Python configuration object, then load data and query it. The same load and query steps are available through a deepsearcher command line interface. For deployment, a FastAPI service started with python main.py exposes a web API at localhost:8000, with interactive docs in the browser. Default modules can be set in a config.yaml file.

Model and Data Source Flexibility

The configuration supports many LLM providers, including OpenAI, DeepSeek, Anthropic Claude, xAI Grok, Gemini, Ollama, Amazon Bedrock and IBM watsonx.ai. Several embedding providers are supported, such as OpenAI, VoyageAI, FastEmbed and Pymilvus built-in models. Vector storage is documented for Milvus (including Milvus Lite and Zilliz Cloud), Qdrant and Azure AI Search. For ingestion, file loaders include Unstructured and Docling, and web crawlers include FireCrawl, Crawl4AI and Jina Reader.

Documented Notes

The README advises that small LLMs may struggle to follow the prompt format and recommends large reasoning models. The documented future plans include better web crawling and support for more vector databases and models.