Overview
People who run a process in a workflow platform deal with its data and a stack of documents together. We added an AI assistant to the process screen itself. A user asks a question in plain language, and the assistant works out which documents, images and step data matter, reads them, and answers with the sources it used, without the user opening each file in turn.
What we built
- Ask about a processQuestions are asked in plain language about one process, with room for a longer description of what the user is after.
- Reads many file typesText is extracted from Word, PDF and PowerPoint files. Images, including pictures embedded in PDFs, are read by a vision model so their content can be asked about too.
- Picks what is relevantA quick first step chooses which documents and images relate to the question, using only their names, so only those are read in full.
- Reads like a person wouldThe assistant is given tools to search a document for a phrase, open a particular page or section, look at an image, read the fields of a step and check who has approved it, and it chooses which to use for each question.
- Search by meaning and keywordDocuments are split into passages and indexed once. Searches combine matching by meaning with matching by keyword, which finds the right passage even when the wording differs.
- Long documentsVery large documents are summarised in parts and then combined, and each document gets a fair share of what the assistant can read, so one large file does not crowd out the rest.
- Sources shownEvery answer lists the documents and pages it used, and the documents are highlighted in the reply.
- Live answersAnswers stream as they are written, with progress messages such as searching or reading a file, and can be stopped at any point.
- Quicker repeat answersExtracted text and search indexes are stored against a fingerprint of each file, so unchanged files are not read again, and old indexes are cleared away automatically.
- Saved conversationsFollow-up questions keep their context. Conversations are saved against the process, reopened, deleted or cleared, with a warning before unsaved work is lost.
- Clear failuresIf the AI service is slow or unavailable the user gets a plain message and the assistant can fall back to a simpler answer, rather than an error screen.
- Part of the existing screensThe assistant sits inside the screens people already use, rather than being a separate tool to open.
