AnswerFetch - Ask your documents anything. Get instant answers.
by•
AnswerFetch turns your scattered docs, training guides, and history into instant answers — so your sales reps close faster, your support team resolves in one reply, and every decision is backed by what your company already knows.
Replies
Hey Mike, congrats on the launch. Been building in the same RAG space for a while at TalkBuildr, the hardest part for us wasn't ingesting the docs, it was dedup on re-ingest and keeping vector search honest when doc content changes. How do you handle that, diff on source_id or just replace the whole collection on re-upload?
@cuygun thx for question. right now the indexed files are shown in the indexed files tab under the admin button. if the file has changed, users can hit the reindex button to update the embeddings and database chunks. it is not currently automatically detected when source files change and is something we are determining about a future upgrade. with sharepoint, and onedrive it works differently than local files and is detected based on the timestamp and updated when a reindex is applied.
@mike_frieswyk Makes sense. Timestamp-based reindex on SharePoint/OneDrive is smart since you get that for free from the source. On local files you'd probably want a content hash per chunk to avoid re-embedding unchanged sections when a user reindexes a big doc, that one's saved us real cost on our side. Solid launch, gl!
@cuygun that would be optimal, but today its just a manual refresh of the file, which is a delete and rechunk on local.
Curious if this works with structured data like TV databases or scheduling files. Building CouchTime and wondering if something like this could help users query their watch history.
@brian_h4 hi Brian, it does the standard sort of ingestion, filtering, and embedding of the chunks based on our chunking algorithm in order to optimize the semantic search abilities the platform offers.
@mike_frieswyk That's helpful. So you're optimizing for semantic search rather than exact matching. Makes sense for natural language queries. For us, the challenge is more about maintaining state across a user's schedule and preferences over time, but the semantic piece could be useful for the "what should I watch" question. Thanks for the context.
@brian_h4 not sure i answered you correctly on prior. i believe it will depending on what the file extensions are for those file types. it has the ability to even look at ocr scanned files and build chunks around what has been scanned.