Context
In many organisations the information exists but cannot be found: it is spread across documents, meeting notes and exchanges accumulated over years. Keyword search is not enough, because the question asked almost never uses the words of the document holding the answer.
The problem
A conversational assistant plugged naively into a corpus raises three immediate problems.
- It can produce a plausible but wrong answer, with nothing to flag it
- It can surface a document to a user who has no right to see it
- It becomes expensive if every question triggers processing of the whole corpus
- It goes stale as soon as documents change, unless the index is kept current
The solution
A retrieval-augmented architecture: find the relevant passages first, only have the synthesis generated, and always cite the documents used.
- Ingestion and chunking of documents into usable passages
- Vector indexing, enabling search on meaning rather than words alone
- Retrieval of relevant passages before any generation
- Natural-language answers, with source documents and a preview alongside
- Access-rights filtering applied to the search, upstream of generation
- Re-indexing of modified documents, so the index follows the corpus
My role
Design of the ingestion and indexing chain, the chunking strategy, the retrieval mechanism, the model integration and the conversation interface, plus the rights filtering.
Constraints
- Access rights must apply to the search, never after the fact on the answer
- Every answer has to be tied to its sources, so it can be verified
- Model call costs to contain, which means sending only the useful passages
- A living corpus: the index has to reflect current documents
- Sometimes confidential documents, which bears on hosting and processing choices
Outcome
An assistant that answers in natural language and shows the documents it relied on, which makes every answer verifiable.
- Source citation makes the tool usable in a professional context
- Rights filtering is applied to the search, and therefore structurally respected
- Cost per question stays controlled, since only relevant passages are sent
Outcomes are described by the capability delivered. No commercial performance metric is claimed here: usage figures belong to the client, and I do not publish numbers I cannot substantiate.