// the find
zilliztech/deep-searcher
Open Source Deep Research Alternative to Reason and Search on Private Data. Written in Python.
DeepSearcher is a RAG framework for running deep research queries over private document collections. You load your docs (PDFs, web pages) into Milvus, then hit it with a question and get a multi-step reasoning loop that searches, reflects, and synthesizes an answer. It's from Zilliz, the company behind Milvus, so the vector DB integration is the strongest part.
The provider abstraction is genuinely well-executed — swapping between OpenAI, DeepSeek, Claude, Gemini, or a local Ollama model is one config line, not a rewrite. The chain-of-rag agent implementation iterates on sub-queries rather than doing a single retrieval pass, which meaningfully improves multi-hop answers. Milvus Lite support means you can run the whole stack locally with a single file URI, no Docker required for prototyping. The evaluation directory with real benchmark data (2WikiMultihopQA) shows someone actually measured whether the thing works.
Hard dependency on Milvus/Zilliz as the primary vector DB — Qdrant and Azure AI Search are listed but feel bolted on, and there's no FAISS or in-memory option for purely ephemeral workloads. Web crawling is still marked under development for most loaders; Firecrawl requires a paid API key, so the 'load from URL' story is weaker than the README implies. The configuration system is a global singleton (`init_config`) which makes it painful to run multiple configurations in the same process or in tests without careful teardown. No streaming support for the query response — for long reasoning chains this means you sit and wait with no feedback.