TECHNOLOGY
Docusense processes documents, turns questionnaires into a structured review, and searches your document library with citations back to the source. This page describes the services that do it. Two further pages go a level deeper into retrieval and into the models themselves.
ARCHITECTURE
Production and sandbox instances have their own backend, PostgreSQL database, vector store, file storage, internal network, and runtime secrets.
A web interface and an API. Every access check is applied in the API rather than in the interface, so it holds regardless of which client makes the request. The API owns questionnaires, review changes, search, processing jobs, and generated PDF responses.
PostgreSQL holds the records, file storage holds the original uploads and generated responses, and the vector store holds the search indexes. The vector store is derived data rather than a source of truth, so it can always be rebuilt from the database.
DOCUMENT PROCESSING
Upload PDF, Word, Excel, or PowerPoint files. Docusense keeps the original and runs it through a fixed sequence of recorded steps, so an administrator can always see where a file is.
A configured parser converts the file into markdown, using optical character recognition and table extraction where the source requires it. Repeated headers and broken markup are removed so that everything downstream reads the same cleaned text.
A questionnaire is split into questions, standalone requirements, and document metadata. Library material is divided along its own headings, and each section keeps the breadcrumb of the headings above it and its position in the original file.
Sections are embedded along with generated summaries and questions, and written to the vector store with references back to the stored source. Re-indexing replaces a document's vectors rather than adding to them.
The retrieval page covers each of these steps in detail, including why the chunking follows document structure and what is generated alongside each section.
MODEL CONFIGURATION
An instance holds a set of configured model servers. Each model is registered for one purpose with its own sampling and prompt settings, and the application asks for the active model of a purpose rather than for a particular vendor or version.
Because routing goes through the purpose, any of the four can be replaced without touching the others, and different purposes can be served by different providers. In a self-hosted deployment all four run in your environment and the core deployment can operate air-gapped.
PROCESSING
Administrators can follow processing status, inspect errors, and retry failed work. Partial results stay visible so a reviewer can decide what to do next.
Status, model, duration, and errors remain available to administrators for each step of each document.
Model work is queued with retries and increasing delays between attempts, and failed work can be queued again manually using the retry settings for that environment.
If only part of a document is extracted, Docusense keeps the available items and shows a warning rather than discarding the result.
IN MORE DEPTH
Claims about AI are only useful if you can check what they mean. These pages set out the mechanics rather than the marketing.
How a document becomes searchable, how meaning and wording are searched together, and how an answer stays tied to a source.
Background on the architecture behind every model in the product, including what it cannot be relied on to do.
