A chat field connected to a vector database can be built quickly. The difficult work begins when an answer must support a real decision.
A document must remain a source
A file, its OCR text, its summary and its retrieval chunks are not the same object. The system must preserve the connection between every excerpt and its original page. Without that link, a citation is only decoration.
Retrieval needs more than one signal
Semantic search understands a general topic well. It may lose a case number, date, name or exact reference. A robust system usually combines semantic retrieval, lexical search and structured filters.
Context must be bounded
Sending every document to the model increases cost and can reduce quality. Explicit, limited and observable context makes it possible to understand why a source was selected.
Uncertainty should be visible
A good interface does not present every response as certain. It shows the sources used, missing documents, contradictions and claims that require human verification.
The final product is therefore not a chatbot. It is a research, evidence and decision system in which the language model is only one component.