Matrix and OpenAI o1: Smarter AI Agents

Matrix and OpenAI o1: Smarter AI Agents

Earlier this year, we launched the first multi-agent AI platform, Matrix.  An “AI associate” for complex work, Matrix can do in seconds the work that used to take an army of associates working around the clock.

When a user types in a prompt, Matrix breaks that complex question into bite-sized analytic steps.  Then the agent gathers and interprets thousands of relevant documents and analyzes text, charts, and tables in their entirety with an infinite effective context window. The final result is returned as a fully verifiable synthesis with citations. If a credit analyst asks Matrix for typical contract terms for a specific client, it can analyze hundreds of credit agreements and extract details such as facilities, term lengths, amortization schedules, call protections, and incremental debt capacities, in an exhaustive, well-formatted analysis.

Since inception, we built Matrix to work with every foundation model, so that our users can always leverage the cutting edge. Its advanced orchestration engine dispatches thousands of LLM calls when a user inputs a complex research query, intelligently routing different tasks to different models based on their specific strengths. 

With the release of o1, Matrix adds OpenAI’s most powerful reasoning model to its tool belt.   

“o1 has set a new standard for the reasoning capabilities of LLMs. When that reasoning power is combined with AI agents and Hebbia’s orchestration, AI can handle even more complex tasks with even more efficiency,” said Shyamal Anadkat, Head of Startup Solutions at OpenAI. 

From our early testing of o1, here are a few examples of work Matrix is even better at:

• Drafting complex documents -  With o1's extended output token capacity, Matrix can now better service users' complex prompts that require long responses. Asking simple questions in Matrix, can now produce a several thousand-word deal memo section.
• Understanding the densest legal and technical documents - o1’s reasoning capabilities enable Matrix to produce exhaustive, well-formatted, and detailed responses when processing complex documents. For example, o1 enabled Matrix to easily identify baskets available under the restricted payments capacity in a credit agreement, with a basic prompt. No former models, even fine-tuned legal models, are as performant. o1 yielded stronger results on 52% of complex prompts on dense Credit Agreements compared to other models
• Multi-step data extraction - Matrix could already extract and read tables and charts, but o1’s reasoning capabilities enable it to think critically and understand the context in which the data appears. This enables o1 to go steps beyond every other model, driving to an answer from context clues alone. For example, Matrix can now infer revenue split concentration across assets in a dense table, even if that split isn’t explicitly mentioned. When testing complex queries across tables, o1 was preferred in 92% of cases.
• Uncovering impacts of market-changing news - News cycles often trigger market changes, but it takes time to gather the necessary information to contextualize and understand the second and third-order impacts of any event. Matrix users are now leveraging o1’s reasoning over Hebbia’s real time data sources to contextualize events and swings into broader thematic relationships they can apply across their portfolios. 

As foundation models continue to improve, the world’s best financial and legal firms will continue to be the first to use them, on Hebbia. 

And Matrix will only get better, continuing to be the platform enterprise users turn to as they define the cutting edge of AI for work.