Open search for agents that beats closed rivals on multi-hop queries

What it does

OpenDeepSearch is a lightweight search and retrieval tool built specifically for AI agents. It fetches web results, reranks them with semantic models like Qwen2-7B-instruct or Jina AI, and feeds the context into agent frameworks such as Hugging Face’s SmolAgents. It offers two modes: a fast default for quick lookups and a slower “Pro” deep-search mode for complex, multi-step queries.

The interesting bit

The project claims to match closed-source search on single-hop benchmarks like SimpleQA while outperforming them on the multi-hop FRAMES benchmark—suggesting its reranking and retrieval pipeline is tuned for reasoning that spans multiple sources. It also integrates via LiteLLM, so you can plug in models from OpenAI, Anthropic, OpenRouter, or self-hosted endpoints without rewriting the tool.

Key highlights

• Uses Crawl4AI and semantic rerankers to dig past surface-level search results
• Two operating modes: Default for speed, Pro Mode for deeper accuracy at higher latency
• Built to drop into SmolAgents and agent frameworks like CodeAgent
• Model-agnostic through LiteLLM, supporting OpenAI, Anthropic, Gemini, HuggingFace, Fireworks, and others
• Search backend is swappable between Serper.dev and self-hosted SearXNG instances

Caveats

• Not batteries-included: you must bring your own search API (Serper.dev or SearXNG), reranker (Jina API or self-hosted Infinity), and LiteLLM-compatible model keys
• Pro Mode’s deeper accuracy comes at the cost of additional processing time

Verdict

Worth a look if you are building agent workflows in Python and need transparent, multi-hop web search that you can tune and host yourself. Skip it if you want a single-vendor, fully managed search API with no assembly required.