Our API takes natural-language prompts like "Nvidia M&A history" and returns visual answers and grounding text sourced from real-time, structured data (example: https://trytako.com/card/YHloo1Ea7GRnBr_s5r6s/).

Most AI systems can’t effectively reason about real-time, structured data. One reason is access: a lot of the most valuable info is trapped in databases web crawlers can't index. Google solves this with a team of 2k+ engineers that ingest data (stocks, sports, etc) into a proprietary Knowledge Graph. Our goal is to offer developers the same knowledge search + visualization primitives Google has built, tailored for AI use cases, and delivered via API.

We seek to augment LLM’s capabilities, and this means that most of our biggest technical challenges stem from not getting a lot for “free” from LLMs. For example, RAG architectures that generate final outputs with LLMs introduce accuracy issues we can’t tolerate, and are too slow. We’ve built a Generative Augmented Search (GAS) architecture that uses LLMs (currently Llama 3.3-70B on Cerebras) to analyze input queries (~200 ms) but use deterministic retrieval for most output generation. The data in our knowledge graph generally isn’t available in LLMs or the web, so we have to acquire it directly from sources (including licensing it from authoritative providers like S&P Global). A limitation of this approach is that some developers want us to offer the flexibility of LLM analysis across web sources, even if it means tolerating non-authoritative sourcing and some hallucination. We’re working on some solutions to that now.

Curious to hear how other people are fighting hallucination.

I'd love your feedback on the product (and happy to discuss/answer questions about it/the tech stack)