Question answering over enterprise documents
Ask your documents. Get the answer with its clause.
Brainer ties every sentence of an answer to a passage in your documents. A sentence it cannot tie, it does not write. Without evidence it says "insufficient information" and stops.
Built for Turkish documents: regulations, contracts, procedures, email. Runs on your own servers or in a deployment we operate.
- Every sentence cited
- Deployable on your own servers
- Measurements published
Built for
- Finance and insurance
- Legal and compliance
- Healthcare
- Public sector and universities
- Manufacturing
The flow
Question, search, source, answer.
A question comes in. The documents are searched, the passage that answers it is found, and the answer is tied to that passage. If no passage is found, no answer is written.
This window is an illustration: the sequence is real, the documents are examples. For real measurements, see the measurement log. Measurement log
The problem
The information exists. Finding it is slow. AI makes it up when it can't.
According to the McKinsey Global Institute's 2012 study, knowledge workers spend 19% of their week searching for and gathering information. General-purpose chat assistants shorten that time and add a new problem: it is not clear which document an answer came from, and an answer arrives even when no document has one.
McKinsey Global Institute, The social economy (2012), Exhibit 2
- ✕ 01Which document and which clause an answer rests on is not shown.
- ✕ 02Information absent from the documents is written as fluently as if it were there.
- ✕ 03Turkish suffixes break search: a query for "sözleşmede" does not find the page that says "sözleşmenin".
- ✕ 04Who may see which document is not part of the search; the permission filter is added afterwards, or not at all.
Each of these four points has a measurable counterpart in Brainer: citations, abstaining, Turkish stemming, and a permission filter inside the search.
Measurement
Our own measurement, weak spot included.
A 39-question Turkish test set: law, finance, insurance. 30 questions are answered by the documents, 9 are not. Measured 15 June 2026.
- 29/29
- Cited answers
- a valid citation on every answer given; ghost citations: 0
- 9/9
- Correct stops
- stopped on all 9 questions the documents do not answer
- 1/29
- Flagged as unsupported
- the one answer the judge model found borderline
- 1/30
- Needless stop
- the question it stopped on although the answer was in the documents
With the production system prompt; the judge is a separate model. This is our own set. In the pilot, the same measurement is repeated on your documents and the report is yours. Method, findings and limits: Measurement log
Contents
- 01How an answer is verifiedCitation check, sentence-level verification, abstaining. With the thresholds.Read
- 02Three differences, and what we don't doWe show what we do well, and we write down what we don't do.Read
- 03Permissions, records, deletionThe permission filter inside the search. Every event in a chained log. Deletion with a signed certificate.Read
- 04Measurement logThe source of every number on this site. Test set, method, findings, weak spots.Read
- 05Engineering notesThe reasoning behind the decisions: tenant isolation, the abstain policy, Turkish search.Read
- 06PricingPer organization; the number of users does not set the price. Two deployment models.Read
Frequently asked
What evaluators ask first
- Do our documents leave our network?
- On your own servers, no: document, question and answer text stay inside your network, and every component including the language model runs locally. In the deployment we operate, documents are held on our infrastructure; for a tenant with sensitive mode on, answers are generated only by the local model and no text goes to an external provider.
- Which sources does it connect to?
- Google Drive, email (Gmail, Microsoft 365, IMAP), S3-compatible object stores and manual upload. PDF, Word, Excel, PowerPoint, email and scanned documents are processed.
- Why does it sometimes say "insufficient information"?
- In three cases: retrieval confidence is below the threshold, the answer's sentences cannot be verified against their source, or a citation is not in the evidence list. This is a design decision; the thresholds are deliberate because a wrong answer is expensive. We track the rate per tenant; the target band in the pilot is 5% to 15%.
- Can users see each other's documents?
- No. The permission filter is applied inside the search engine, and the returned list is checked a second time in the database. Separation between organizations has three independent layers and is verified by an automated test on every release; the last full run: 2,010 queries, 0 leaks.
- How long does setup take?
- We plan 2 weeks for the pilot: the first week connecting sources and indexing, the second measuring with your questions. On your own servers, hardware and network preparation on your side takes additional time.
- How is the price set?
- Per organization: document volume, number of departments and deployment model. The number of users does not enter the price. The quote comes at the end of the pilot, together with the measurement results.
Built for
Where the document is binding
Thousands of documents, several departments, an audit obligation.
Finance and insurance
Legislation, product rules, policy terms. An answer cannot be used without its clause.
Legal and compliance
Contract templates, internal regulations, board decisions. Clause-level citation is required.
Healthcare
Protocols, procedures, personal health data. Masking and deployment location decide.
Public sector and municipalities
Regulations, circulars, correspondence. Air-gapped deployment and a full audit log.
Universities
Directives, senate decisions, administrative procedures. Multi-unit, permission-aware access.
Manufacturing
Quality documents, work instructions, maintenance records. Turkish technical term search.
Measure it on your own documents.
A two-week pilot: your sources are connected, the result is measured with your questions, and the report is yours.