Open Deep Search arrives to problem Perplexity and ChatGPT Search

Researchers at Sentient Basis have launched Open Deep Search (ODS), an open-source framework that may match the standard of proprietary AI search options resembling Perplexity and ChatGPT Search. ODS equips giant language fashions (LLMs) with superior reasoning brokers that may use internet search and different instruments to reply questions. 

For enterprises on the lookout for customizable AI search instruments, ODS affords a compelling, high-performance different to closed industrial options.

The AI search panorama

Fashionable AI search instruments like Perplexity and ChatGPT Search can present up-to-date solutions by combining LLMs’ information and reasoning capabilities with internet search. Nevertheless, these options are usually proprietary and closed-source, making it troublesome to customise them and undertake them for particular purposes. 

“Most innovation in AI search has happened behind closed doors. Open-source efforts have historically lagged in usability and performance,” Himanshu Tyagi, co-founder of Sentient, advised VentureBeat. “ODS aims to close that gap, showing that open systems can compete with, and even surpass, closed counterparts on quality, speed, and flexibility.”

Open Deep Search (ODS) structure

Open Deep Search (ODS) is designed as a plug-and-play system that may be built-in with open-source fashions like DeepSeek-R1 and closed fashions like GPT-4o and Claude.

ODS contains two core parts, each leveraging the chosen base LLM:

Open Search Instrument: This part takes a question and retrieves data from the online that may be given to the LLM as context. The open Search Instrument performs a number of key actions to enhance search outcomes and be certain that it supplies related context to the mannequin. First, it rephrases the unique question in numerous methods to broaden the search protection and seize numerous views. The device then fetches outcomes from a search engine, extracts context from the highest outcomes (snippets and linked pages), and applies chunking and re-ranking methods to filter for essentially the most related content material. It additionally has customized dealing with for particular sources like Wikipedia, ArXiv and PubMed, and will be prompted to prioritize dependable sources when encountering conflicting data.

Open Reasoning Agent: This agent receives the consumer’s question and makes use of the bottom LLM and numerous instruments (together with the Open Search Instrument) to formulate a closing reply. Sentient supplies two distinct agent architectures inside ODS:

ODS-v1: This model employs a ReAct agent framework mixed with Chain-of-Thought (CoT) reasoning. ReAct brokers interleave reasoning steps (“thoughts”) with actions (like utilizing the search device) and observations (the outcomes of instruments). ODS-v1 makes use of ReAct iteratively to reach at a solution. If the ReAct agent struggles (as decided by a separate decide mannequin), it defaults to a CoT Self-Consistency, which samples a number of CoT responses from the mannequin and makes use of the reply that exhibits up most frequently.

ODS-v2: This model leverages Chain-of-Code (CoC) and a CodeAct agent, applied utilizing the Hugging Face SmolAgents library. CoC makes use of the LLM’s means to generate and execute code snippets to resolve issues, whereas CodeAct makes use of code technology for planning actions. ODS-v2 can orchestrate a number of instruments and brokers, permitting it to sort out extra complicated duties which will require subtle planning and doubtlessly a number of search iterations.

“While tools like ChatGPT or Grok offer ‘deep research’ via conversational agents, ODS operates at a different layer—more akin to the infrastructure behind Perplexity AI—providing the underlying architecture that powers intelligent retrieval, not just summaries,” Tyagi stated.

Efficiency and sensible outcomes

Sentient evaluated ODS by pairing it with the open-source DeepSeek-R1 mannequin and testing it towards standard closed-source rivals like Perplexity AI and OpenAI’s GPT-4o Search Preview, in addition to standalone LLMs like GPT-4o and Llama-3.1-70B. They used the FRAMES and SimpleQA question-answering benchmarks, adapting them to guage the accuracy of search-enabled AI programs.

The outcomes exhibit ODS’s competitiveness. Each ODS-v1 and ODS-v2, when mixed with DeepSeek-R1, outperformed Perplexity’s flagship merchandise. Notably, ODS-v2 paired with DeepSeek-R1 surpassed the GPT-4o Search Preview on the complicated FRAMES benchmark and practically matched it on SimpleQA.

An attention-grabbing statement was the framework’s effectivity. The reasoning brokers in each ODS variations discovered to make use of the search device judiciously, usually deciding whether or not a further search was mandatory primarily based on the standard of the preliminary outcomes. As an example, ODS-v2 used fewer internet searches on the less complicated SimpleQA duties in comparison with the extra complicated, multi-hop queries in FRAMES, optimizing useful resource consumption.

Implications for the enterprise

For enterprises in search of highly effective AI reasoning capabilities grounded in real-time data, ODS presents a promising answer that gives a clear, customizable and high-performing different to proprietary AI search programs. The power to plug in most well-liked open-source LLMs and instruments provides organizations larger management over their AI stack and avoids vendor lock-in.

“ODS was built with modularity in mind,” Tyagi stated. “It selects which tools to use dynamically, based on descriptions provided in the prompt. This means it can interact with unfamiliar tools fluently—as long as they’re well-described—without requiring prior exposure.”

Nevertheless, he acknowledged that ODS efficiency can degrade when the toolset turns into bloated, “so careful design matters.”

Sentient has launched the code for ODS on GitHub.

“Initially, the strength of Perplexity and ChatGPT was their advanced technology, but with ODS, we’ve leveled this technological playing field,” Tyagi stated. “We now aim to surpass their capabilities through our ‘open inputs and open outputs’ strategy, enabling users to seamlessly integrate custom agents into Sentient Chat.”