Ask questions in Arabic or English across every document you hold. Each answer carries citations back to page and coordinates, and the language model that generates it runs on your own GPUs — so the prompt, the passages and the answer never leave the building.
Answer quality on your own corpus is measured during the Proof of Value against a human-adjudicated question set you help write, including questions the documents cannot answer, so refusal behaviour is tested rather than assumed.
A confident answer with no source is worse than no answer, because it costs a person an hour to disprove and a regulator one question to reject.
Every claim in an answer resolves to a bounding box on a specific page of a specific document, so a reviewer can open the source and see the sentence the answer came from.
When retrieval finds nothing that supports an answer, the response says so and names what was searched. The system is tuned to prefer a refusal over a fluent guess.
Ask in Arabic about English documents, or in English about Arabic documents. Retrieval crosses scripts, and citations point at the source region in its original script.
Access control is applied to the index before generation, using the identity from your own OIDC or SAML provider. The model is never shown a passage the user could not open themselves.
Open-weight models ship in the deployment bundle and are served inside your network. There is no hosted inference endpoint and no overflow call to a cloud model under load.
Retrieval draws on the analysed document — table grids, typed fields and the entity graph — so a question about a total finds the cell, not a paragraph that happens to mention totals.
Most private Q&A projects fail at the same place: the corpus was flattened into text before it was understood, so the index has no idea which number was in which cell, which page a clause spanned, or that three documents describe one person. DocxIntel asks questions of the output of five other capabilities.
The quality of an answer is decided long before the question is asked.
An adjuster asks: was this vehicle covered on the incident date, and does the repair estimate exceed the excess?
Your OIDC or SAML provider supplies the identity. Roles and attributes resolve into a retrieval filter before any search runs, so the searchable corpus for this question is only what this person may see.
Hybrid lexical and vector search across passages, table cells, typed fields and the entity graph, in both Arabic and English, returns candidate evidence with document, page and coordinates attached.
Candidates are re-ranked, and the set is tested against the question. If nothing supports an answer — the policy schedule for that date is simply not in the estate — the pipeline stops here and returns a refusal.
The language model runs on your GPUs. It receives the retrieved passages and the question, and is constrained to answer from them. No prompt, passage or answer crosses your network boundary.
Claims that cannot be attributed to a retrieved region are dropped rather than shown. The answer arrives as text plus a citation list of documents, pages and bounding boxes.
Question, user, retrieval filter, evidence used, answer and citations are appended to your own audit log and can be shipped to your SIEM — so an answer given in March is reviewable in November.
Cloud models are excellent. The problem is the shape of the transaction: to get an answer about a customer file, the customer file has to be sent to somebody else.
Where prompts and documents go
Works with no internet connection
Granularity of citations
Behaviour with no supporting evidence
Permissions applied to retrieval
Retention of submitted content
Training on your content
Cost of asking ten thousand questions
Arabic questions over Arabic documents
Comparison reflects publicly documented behaviour of managed LLM, parsing and retrieval services as of 2026, including per-token and per-page metering, default caching of parsed data and self-hosting gated to enterprise plans. Verify against current vendor documentation before making a decision.
What runs where, in a standard deployment, when the Ask capability is enabled.
Model selection, GPU sizing and concurrency targets for a specific deployment are agreed in writing during the Proof of Value, against your own corpus and your own question set.
What gets asked when the phrase generative AI appears in a paper going to a board committee.
Two things. Answers are generated only from retrieved passages, and every claim in the answer must be attributable to a specific region of a specific page or it is not emitted. When the retrieved evidence does not support an answer, the response is an explicit refusal naming what was searched and what was missing, rather than a plausible sentence. A refusal is a correct answer, and the system is tuned to prefer it over a guess.
