Can I keep my OpenAI models and use Hippocampus for retrieval?

Yes: Hippocampus works with any language model, OpenAI's included. You choose which model answers, set per workspace; on Corporate you can use your own model key or let an admin choose the models. Retrieval and citations stay in Hippocampus whichever model answers.

Does OpenAI File Search give page-level citations?

Page-level citation is not documented for File Search: answers carry a file citation with the file ID, the file name and a character index (checked October 2026). Hippocampus returns the document, section path, start and end page, byte offsets and the text used for every source.

Can File Search restrict results per user?

Not as identity-based access control: the documented route is file attributes and metadata filters that the caller supplies (checked October 2026). Hippocampus enforces document permissions inside the search, before ranking, and returns nothing if permissions cannot be checked.

Can I keep my documents in the EU?

Both store documents in Europe. OpenAI lists regional storage in Europe and regional processing in Europe for File Search in the Responses API, with approval needed for non-US regions (checked October 2026). Hippocampus stores documents and the knowledge structure in the EU (Finland); model inference runs in the US by default and can run EU-hosted or in your own environment on Corporate.

Is my data used to train models?

Neither uses it by default. OpenAI does not train on API data unless you opt in (checked October 2026). Hippocampus never uses customer data to train AI, and its contracts forbid the model providers from doing so.

Comparison

Hippocampus vs OpenAI File Search

18 October 2026

Last updated: October 2026

By Andreas Dahlström, Head of Deeptech Team, Co-Founder

Hippocampus is the better choice when every answer must be checked against a page and when people and agents may only see what they are allowed to read. OpenAI File Search is the better choice when you already build on OpenAI's Responses API, want retrieval as one more hosted tool, want to tune hybrid search yourself, or need fast answers. On our FDA benchmark, with GPT-5.4 mini answering for both, Hippocampus answered 40 of 48 questions correctly and File Search 36 of 48*, a gap within run-to-run variation. File Search was faster: a median of 7 seconds per answer against 20 for Hippocampus.

This article compares the two products dimension by dimension. The benchmark run itself, with five systems side by side and the questions where they differ, is our earlier post Hippocampus against OpenAI File Search and PageIndex; this article does not repeat it. We also compare Hippocampus vs PageIndex and Hippocampus vs Vertex AI Search.

Summary

Dimension	Hippocampus	OpenAI File Search
Hosting region and self-hosting	Documents and the knowledge structure are stored in the EU (Finland), encrypted. Model inference runs in the US by default; on Corporate it can run EU-hosted or in your own environment. Own cloud or on-premises hosting is part of the Enterprise package. Security	Data residency is set per project, with regional storage in the US, Europe and eight other regions. Vector stores have no regional processing; File Search in the Responses API has regional processing in the US and Europe only. Non-US regions need approval. Self-hosting: not documented. Your data (checked October 2026)
Source attribution (page-level citation)	Every answer carries the document, section path, start and end page, byte offsets and the text used. Citations come from the retrieval step, not from the model.	Answers carry `file_citation` annotations with the file ID, the file name and a character index. Page-level citation: not documented. File Search guide (checked October 2026)
Permission enforcement at query time	Enforced inside the search as a filter before ranking; returns nothing if permissions cannot be checked. A summary over several documents is visible only to people who can read all of them. Agents obey the same permissions. Security	Per-user access control: not documented. The caller can narrow the files searched with file attributes and metadata filters. Retrieval guide (checked October 2026)
Behaviour on unanswerable questions	When retrieval comes back empty the question is refused, not guessed. A refusal costs the same credits as an answer.	Refusal behaviour: not documented. A score threshold and a result limit reduce the chunks returned. File Search guide, Retrieval guide (checked October 2026)
Retrieval architecture (structured memory vs vector index vs file search)	Structured memory built at import: a tree per document with an index summary, section summaries and leaf sections with pages and byte ranges. Two funnels, top-down over the summaries and bottom-up keyword plus embedding search over leaf sections, fused into one ranking. Developers	Vector store: files are chunked (800 tokens with 400 overlap by default), embedded and indexed on upload. Semantic search with optional query rewriting and weighted hybrid keyword search, up to 50 results. Retrieval guide (checked October 2026)
Supported answering models	Any language model. You choose which model answers, set per workspace; on Corporate you can use your own model key or have administrators choose the models. Developers	Runs as a tool in OpenAI's Responses API. A list of supported models: not documented. File Search guide (checked October 2026)
API and MCP availability	REST API and an MCP server. The MCP server is in beta; OAuth sign-in on the shared cloud is being completed. The API on the Free tier is in beta. Developers	Configured as a tool in the Responses API. The Responses API can call remote MCP servers; File Search as a hosted MCP server: not documented. File Search guide, MCP guide (checked October 2026)
Data export format	Open Knowledge Format: a folder per document of Markdown with YAML front matter, holding the index summary, the full text, the original file and a folder per section. Readable through the API, exportable while the account is active. Developers	An endpoint returns the parsed text of a vector store file, paginated. Export of embeddings or of a whole store: not documented. Vector store file content (checked October 2026)
Use of customer data for training	Never used to train AI, not by Thalius and not by the model providers; their contracts forbid it. Security	API data is not used for training by default; sharing it requires an explicit opt-in. Your data (checked October 2026)
Pricing model and free tier	Credits, one unit on every tier. Free tier with monthly included credits and no card, Starter, and Corporate by quote. A refusal costs the same as an answer. Amounts on Pricing	Usage-based: vector storage per GB per day with a free allowance, plus a fee per File Search call. Amounts on the pricing page (checked October 2026)

Benchmark

Ten public FDA documents for Saxenda, Wegovy and Zepbound (2,198 pages), sixteen questions written by Atrogi's team, each asked three times in independent conversations. Every system answered through the same model, GPT-5.4 mini, on 17 September 2026. A grading model, GPT-5 with three votes and a majority verdict, compared each answer with a reference answer; declining the control question counts as correct.

Hippocampus: 40 of 48 correct*, 79% weighted, median 20 seconds per answer.
OpenAI File Search: 36 of 48 correct*, the average of two runs that scored 38 and 34, 64% weighted, median 7 seconds per answer.

The weighted score counts easy questions once, medium twice and hard three times. On the control question, which the documents cannot answer, File Search gave a ranking as its answer in four of six attempts; Hippocampus said each time that the documents cannot support a ranking, and in two of three answers offered the approval order as a labelled proxy.

What this result does not show:

The gap between Hippocampus and File Search is within run-to-run variation: File Search's two identical runs differed by four answers.
Hippocampus configurations tested on 16 to 17 September scored between 37 and 40 of 48. The figure above is a run that scored 40.
The search guide Hippocampus gives the model was refined on these same sixteen questions. The other systems ran on vendor defaults plus a one-sentence instruction to answer only from the material.
File Search ran twice. The other systems ran once, so differences of a few answers can come from variation alone.
Hippocampus is the slowest system tested: a median of 20 seconds per answer against 5 to 8 seconds for the others.
The grading model is OpenAI's GPT-5, and the answering model and two of the systems tested are also OpenAI's.

The full table with all five systems is in Hippocampus against OpenAI File Search and PageIndex. The corpus, the questions and the results with three answering models are in the FDA GLP-1 benchmark, and every published result is on Evidence.

* Answering model GPT-5.4 mini for every system. Grading by GPT-5, three votes, majority verdict. Thalius evaluation, 17 September 2026.

Where OpenAI File Search is the better choice

You already build on OpenAI's Responses API and want retrieval as one more tool in the same call: files are chunked, embedded and indexed on upload, with no separate service to run (Retrieval guide, checked October 2026).
You want to tune retrieval yourself: hybrid search weights, query rewriting, a score threshold and attribute filters are all exposed (Retrieval guide, checked October 2026).
You need storage in a region outside the EU: OpenAI lists regional storage in ten regions, among them Australia, Canada, Japan and the UAE, with regional processing in the US and Europe for File Search (Your data, checked October 2026).
Speed matters more than depth: in our FDA benchmark File Search answered in a median of 7 seconds against 20 for Hippocampus.

Try Hippocampus on the same documents

Upload the documents you would put in a File Search vector store and ask Hippocampus the same questions, with an OpenAI model answering if you like. Start free, no card. Before you do, read how Hippocampus enforces permissions and where it stores documents, the API, the MCP server and the export format, how credits work on each tier and own-cloud or on-premises hosting on the Enterprise package.

Sources and last updated

Last updated: October 2026.