Wrong chunk at scale
Your RAG demo worked on ten documents. At ten thousand it returns the wrong chunk.
POST /search
POST /search query: How does a customer reset their password?
passage: Open Settings, then Account, and choose Reset password. Help centre guide / Account / p. 4
passage: The customer gets a reset link by email. Help centre guide / Account / p. 4
POST /search query: What is on next quarter's roadmap?
refusal: the documents do not cover this
What it does
Upload files or connect a source. Hippocampus turns each document into a tree of sections with summaries.
Call POST /ask or POST /search, or connect your agent to the MCP server.
The passage, the document and the page, or a refusal if no passage supports an answer.
Hippocampus refuses the question if your documents hold no answer.
The problem
Your RAG demo worked on ten documents. At ten thousand it returns the wrong chunk.
You spent a month on chunk sizes and rerankers and the eval score stayed where it was.
Your agent reads whole documents to find one paragraph, and you pay for each token.
What you can count on
On FinanceBench, Hippocampus scores higher than Azure AI Search, Vertex AI Search and OpenAI File Search.
Each passage comes with its document, section path and page range.
If no passage supports an answer, you get a refusal. A refusal is your signal to hand over to a person.
Three endpoints, or one MCP server for coding agents and chat assistants.
Benchmark
Basic vector search is cheap per query and misses more answers. Hippocampus costs more per query than that and gets more answers right.
Speed is the trade. Median search time is 1.8 s. Azure AI Search answers in 0.4 s and Vertex AI Search in 0.3 s, with lower scores.
FinanceBench: 150 analyst questions on 84 US company filings. Hippocampus 92.2%, Azure AI Search 76.6%, OpenAI File Search 52.3%, Vertex AI Search 40.6%. GPT-5.4 mini wrote every answer.
Where it sits in your stack
Your agent asks. Hippocampus answers from your documents, with the page, or says it cannot.
A refusal is a signal the agent can act on: ask the user, hand over to a person or stop.
Each passage comes with its document, section path and page range.
It reads a document into the memory.
The MCP server exposes ask, search and ingest to agents.
Select a part of the diagram to see its flow.
Hippocampus turns each document into a tree of sections with summaries. A question runs two searches over that tree: top down through the summaries, and bottom up with keywords and embeddings on the leaf sections.
How much notice do we need to give?
Model
writes the answer
Three months, in writing.
Sent to the registered address.
Click a node to read its summary.
2 of 6 passages went to the model. The document did not.
Hippocampus merges the two rankings. Retrieval uses no generative model. The model you choose writes the answer from the passages.
Example with a fictional document.
Fits your setup
Ask a model about a few pages and it does well. Ask it across thousands and its answers get worse. Hippocampus keeps the model on the right pages, so it stays accurate as the collection grows.
Connects to SharePoint, Google Drive and Box.
PDF, Word, PowerPoint, Excel and scanned pages.
EU-hosted by default, or in your own cloud or on your own servers.
You set who sees what by role or function.
You choose the model per workspace and can change it without importing again.
Pick another model. The memory stays as it is.
Questions
The Free plan holds 2,000 pages, needs no card and includes the API and the MCP server. The Starter plan is EUR 30 a month for 25,000 pages. The Corporate plan has custom pricing.
Median search time is 1.8 s.
Yes. POST /search returns passages only.
Hippocampus refuses the question. A refusal costs the same as an answer.
Yes. You export each document as a Markdown folder in Open Knowledge Format.