An open-source rival to Perplexity and ChatGPT Search

Open Deep Search (ODS) is an open-source framework that lets you build a full-fledged AI search engine capable of competing with Perplexity and ChatGPT Search, right on top of models like DeepSeek-R1 or GPT-4o.

Most modern AI search innovations happen behind the closed doors of proprietary products, making deep customization tough for developers and teams. Sentient Foundation built ODS to fix that, offering a plug-and-play architecture that treats intelligent retrieval as an infrastructure layer rather than a locked-down SaaS feature.

How it actually works

ODS splits the heavy lifting into two distinct components: the Open Search Tool and the Open Reasoning Agent.

The Open Search Tool handles the information retrieval. Instead of just firing off your original query, it rephrases your prompt into several variations to capture different perspectives. It grabs results from a search engine, chunks and re-ranks content from snippets and full web pages, and applies specialized handling for sources like Wikipedia, arXiv, and PubMed to prioritize trustworthy data when facts clash.

Sitting on top of that is the Open Reasoning Agent, which drives the thinking process. Sentient ships ODS with two separate agent flavors.

ODS-v1 uses a standard ReAct framework paired with Chain-of-Thought reasoning, constantly interleaving thoughts, tool actions, and observations. If the agent gets stuck, a separate judge model steps in to trigger a self-consistency check, sampling multiple reasoning paths to find the most frequent answer.

ODS-v2 takes things further by adopting CodeAct and Chain-of-Code via Hugging Face's SmolAgents library. Instead of just generating text commands, it writes and executes code snippets to plan actions and orchestrate multiple tools, making it much better at handling messy, multi-step queries.

The benchmark numbers

ODS is fully model-agnostic. You can hook it up to open weights like DeepSeek-R1 or closed APIs like GPT-4o and Claude.

Sentient tested the setup on the FRAMES and SimpleQA benchmarks. When paired with DeepSeek-R1, both ODS-v1 and ODS-v2 outperformed Perplexity's flagship offerings.

On the complex FRAMES benchmark, which requires multi-hop reasoning across multiple sources, ODS-v2 with DeepSeek-R1 actually beat OpenAI's GPT-4o Search Preview, while nearly matching it on SimpleQA.

Just as important, the agents proved surprisingly efficient. Rather than spamming web searches on every turn, the system learned to adapt its query frequency based on the difficulty of the prompt, saving searches on simple questions and going deep only when necessary.

Why it matters

If you want to deploy a private, custom deep-research engine for your company without funneling sensitive queries into closed commercial platforms, ODS is probably the most capable open blueprint available right now.

It closes a massive gap in open-source AI. Until now, matching Perplexity-level retrieval required cobbling together brittle DIY RAG pipelines; ODS shows that an open model like DeepSeek-R1 with clever agent scaffolding can stand toe-to-toe with frontier search products. Just keep in mind that ODS is the orchestrator: your actual performance will still depend heavily on the web search APIs you connect and how much effort you put into system tuning.