What is document intelligence?
Document intelligence is software that reads a pile of documents, works out who and what they mention, and answers questions with a source. It differs from a keyword search because it knows that two spellings of a name are the same person and that a fact may have been true only for a period.
Restrologic built and runs such a system for a client we do not name. The build is summarised in the document intelligence case study. This page explains the problem and the approach for anyone who handles many documents.
Who has this problem?
Anyone who answers questions from a large set of files has this problem. In agencies, legal teams, finance teams and research groups, the answer exists somewhere in the archive and nobody knows which file.
The pattern is familiar:
- A senior person asks “what did we say to this client about that?” and a junior spends an afternoon searching.
- The same person appears as “R. Sharma”, “Rakesh Sharma” and “Mr Sharma” in different files.
- A scanned contract or a chart inside a slide deck cannot be searched at all.
- Two documents disagree, and nobody can tell which is newer.
- A summary from a chatbot sounds right but cannot say where it came from.
How does an AI answer questions from your documents?
An AI answers questions from your documents by finding the relevant passages and facts first, and then writing an answer from only those. Every statement in the answer points back to a file and a page.
The citation matters more than the wording. A lawyer or an analyst cannot act on an answer they cannot verify. When the reply says “see the 2024 agreement, page 7”, the reader checks it in seconds and trusts the rest.
Why do facts need dates?
Facts need dates because most things in documents change. A person changes role, a price is revised and a contract is renewed on new terms. An answer that ignores dates is wrong half the time.
In the system we built, each fact has a time attached. If one file says a person was a manager in 2022 and another says they became a director from March 2025, both are kept, each with its dates. Ask about 2023 and you get the manager role. Ask about today and you get the director role. We wrote about the method in document question answering with dated facts.
How does it read scanned pages and charts?
It reads scanned pages and charts with a vision model, which looks at the page image and extracts the text, tables and numbers on it. This turns a photographed contract or a bar chart in a deck into facts that can be searched.
Reading images costs more than reading text, so the system uses the vision model only where a page needs it. Plain text files take the cheaper route. The extracted facts join the same knowledge graph, so one question can draw on a typed memo and a scanned letter together.
How does it handle names that vary?
It handles varied names by merging variants that refer to the same entity. The same firm may appear with and without “Pvt Ltd”, and a person may appear with or without an initial.
A wrong merge joins two different people and corrupts every answer about them, so merging is a step worth getting right. The knowledge graph then shows one node per real entity, with every mention linked beneath it.
| Without merging | With merging |
|---|---|
| Three entries for one person | One entry with three spellings |
| Facts split across entries | Facts in one place |
| Search misses variants | Search finds all of them |
What does it cost to run?
Most questions cost about $0.001 to answer. The cost stays low because the system narrows the search to a small set of facts and passages before it calls a model, so the model reads a little text and not the whole archive.
Setup costs more, since reading and structuring every document is done once up front and again for new files. We estimate both parts before we start, using a sample of your documents.
How safe is it to use with sensitive files?
Our AI agents run any generated code in a throwaway sandbox with no network access, and their database access is checked so it can only read. Tell us how sensitive your files are and we will agree where the data lives before we start.
Model choice is per client, since some teams prefer one provider for policy reasons. We can work with Claude, GPT, DeepSeek or Qwen through OpenRouter. The AI agents service page covers how we build and test them.
How would you start?
You start by sending a sample of about fifty documents, including the ugliest scans, and ten questions you wish you could ask. We build a small version, answer your questions with citations, and show you where it fails.
That test tells you more than any demo. If it works, we scale it to the full archive. If you also want an assistant for your numbers, see Reckoner. To begin, contact Deepak through the contact page.