Deep Research for Documents You Actually Own

What it does

DeepSearcher is a Python framework that ingests local documents, indexes them in a vector database—usually Milvus or Zilliz Cloud—and answers natural-language queries by retrieving relevant chunks and synthesizing them with an LLM. It is pitched as an open-source alternative to “deep research” tools, but aimed at enterprise knowledge bases rather than the open web. You can optionally augment results with web crawling, though the core focus is keeping sensitive data in-house.

The interesting bit

The project is fundamentally integration glue, yet it is remarkably vendor-agnostic: it supports well over a dozen LLM providers, from the usual suspects like OpenAI and Anthropic to regional hosts and self-hosted Ollama. That breadth signals Zilliz’s real goal—make the vector database underneath indispensable while letting you bring whatever model you already have an API key for.

Key highlights

• Vector storage via Milvus or Zilliz Cloud, with support for data partitioning to speed retrieval.
• A sprawling list of supported LLMs including DeepSeek, Grok, Claude, Llama 4, Qwen3, Gemini, Bedrock, and Watsonx, swapped via a single configuration call.
• Local file loading built in; web crawling is available but flagged as under development.
• Explicitly designed around data privacy, targeting internal knowledge management and Q&A.

Caveats

• Web crawling capabilities are noted as “under development,” so live internet augmentation may be uneven.
• The README reads like a configuration catalog; how the internal reasoning, search, and evaluation loop actually coordinates is left vague beyond high-level diagrams.

Verdict

Try it if you want an open-source deep-research scaffold over internal documents and refuse to lock yourself into one LLM vendor. Pass if you need a finished end-user product rather than a framework to wire up.