Stay on top of your data. Reduce risk. Create value.
Document intelligence for investment funds and finance teams. VectoGrid reads your documents, verifies every figure against its source, and turns them into data you can trust.
90% on FinanceBench, 150 real questions over 368 company reports: the best published result we know of with open-weight models.
Verified, not just cited
Every value is checked against its source. Uncertain values are flagged, never guessed.
Runs on your infrastructure
Fully on-prem if you want it. In fully local mode, nothing leaves your machine.
Built for finance
Portfolio calculations and workflows a general tool does not do.
Why not a chat-with-your-documents tool?
Chat tools answer questions and cite sources. They do not verify the number, compute across documents, or run on your own machine. VectoGrid verifies the value, computes totals across documents, runs fully on-prem and keeps a reusable, structured dataset.
Articles
- How we trained a reranker small enough for a laptop A $9 reranker small enough for a laptop matched the hosted GPU model in our agent: 132 of 150 questions against 133.
- A smaller vector that finds more Qwen3-Embedding-0.6B at 32 bytes a vector beat bge-m3 at full length, and the exact wording of its query prompt was worth 25 answer pages.
- What is the biggest RAG ever built? The largest retrieval systems, why RETRO-style models faded, and why bigger retrieval still hallucinates.
- Is RAG really dead? Agentic Grep came close to Agentic RAG when the question named its document, and trailed it when the answer was spread across a library.
- Should you cap a vector shard at 100,000 vectors? Smaller shards find slightly more of the nearest vectors, and no more of the pages people marked as the answer.
- VectoGrid cuts FinanceBench cost and run time 6x, at 91% Same test, a seventh of the cost: 91%, about 2 cents and 12 seconds a question.
- Document AI over a billion documents: how we would build it Scope before search, keywords and compressed vectors, a reranker per question, enrichment without generative models.
- Is RAG worth the trouble over Elasticsearch? On questions written by people, Elasticsearch loses 1 in 7, standard RAG 1 in 20, VectoGrid none.
- From one report to millions: how VectoGrid prepares document search for growing libraries How VectoGrid prepares document search for growing libraries: half precision, 1-bit vectors, clusters.
- Jev, measured: is it 190× faster and 440× cheaper? Jev against four language models on the same 300 passages, as an extractor and as a checker.
- VectoGrid scores 90% on FinanceBench with open-weight models 150 questions over 368 company reports, every answer and its source published.