I’m consulting for company that is building a solution to connect and ingest data from company websites and documents. We have found that low cost LLM models are often good enough to answer questions accurately about the data. It’s the context of the data (categories, purpose, origin, age) that affects the accuracy of the answers. Some of the challenging issues are determing the source of truth from similar data.
We are curious how others are approaching this problem, i.e. a connecting to a mixture of structured and free form text. Also, if Cdata has a solution or recommendation for this.
A pattern that works reasonably well: keep the structured record as the backbone and treat free-form text as an extraction source that feeds into it, rather than querying both side by side at answer time. At ingest, stamp every chunk with provenance (URL, document type, retrieval date, author/owner) and a confidence or precedence rule per field, then resolve duplicates through entity matching on stable keys before the LLM ever sees the data. That way conflicting values get settled by lineage and rule, not by the model. The LLM’s job narrows to extracting fields from documents and drafting answers over already-reconciled records. Worth checking whether CData’s connectors can carry that metadata through, since that’s usually where it gets lost.
This website stores cookies on your computer. These cookies are used to collect information about how you interact with our website and allow us to remember you. We use this information in order to improve and customize your browsing experience and for analytics and metrics about our visitors both on this website and other media.
Review section III of our Privacy Policy for more information on cookies.
To opt out of “sale” or “sharing” of non-cookie data, visit our Privacy Rights Request Form.