Context

The value of an AI assistant at work grows with how much of the company it can read. So does the risk. Pasting a client contract, a pricing model or a design file into a hosted chatbot sends it to someone else's servers, under someone else's terms. Many firms have, sensibly, banned it outright, and their staff go back to searching folder names and asking whoever has been there longest.

Context is designed around that constraint rather than around it being waived. The models, the index and every query stay inside the company's network, and the system has no outbound route to anything.

Answer quality depends almost entirely on retrieval. Context indexes documents in place and splits them along their real structure, such as sections, clauses, table rows and slide notes, rather than fixed-size chunks. That way a retrieved passage is a coherent unit and not half of two paragraphs.

Each query runs two searches at once. Keyword search catches exact part numbers, clause references and names. Vector search catches the same idea phrased differently. The combined candidates are rescored by a reranking model, and only the strongest passages reach the language model, which is instructed to answer from them alone and to say so when they don't contain an answer.

An internal search tool that ignores access control is a data leak with a search box. Context reads the permissions on each source, including file-share ACLs, wiki spaces and group memberships, and filters at retrieval time. A passage the user couldn't open directly can never be used to answer their question, and can't be summarised around either.

Every answer is shown with numbered citations that open the source at the exact passage. When the documents disagree, Context says so and shows both, rather than choosing one silently.

Context uses current open-weight models, quantised to run well on a single workstation or server GPU, so a mid-sized firm can run it without a data centre. The index updates incrementally as files change. Models can be swapped as better open-weight releases appear, with no change to the rest of the system and no new vendor to approve.

Ingestion, hybrid retrieval with reranking, permission filtering and cited answers are working end to end in the lab against large mixed document sets. Current work is on connectors for common file stores and wikis, and on evaluation: a fixed set of real questions with known answers, scored on every change, so retrieval quality is measured rather than assumed.

Next is handling drawings and scanned documents as first-class sources, so a question about a detail on a drawing can be answered from the drawing itself.