Ask a general AI chatbot a question about a document you uploaded, and you get a confident, well-written answer. The problem shows up when you actually need to trust that answer for something that matters - there's no way to check where it came from without re-reading the whole document yourself, which defeats the point of asking in the first place.
Semantic search finds meaning, not just keywords
Keyword search misses questions phrased differently from the source text. EnergyFile indexes documents with vector embeddings instead - each chunk of a PDF, spreadsheet, or presentation gets encoded so a question can match a passage even if it uses none of the same words. That's what makes "what's our exposure on the northern pipeline" find a paragraph that never uses the word "exposure."
For a larger archive the problem starts earlier: a multi-thousand document archive nobody can search is effectively unusable, however well it's stored.
An answer with no citation is just a guess with good grammar
The part that actually matters: every answer comes with a page-level citation pointing back to the source document. You don't take the AI's word for it - you click through and read the exact paragraph it pulled from. If the citation doesn't say what the answer claims, you know immediately, instead of finding out three decisions later.
The flip side matters as much: when the documents don't cover a question, the assistant should say it doesn't know rather than invent an answer.
OCR means scanned PDFs count too
A lot of research libraries are half searchable text and half scanned pages that never got OCR'd, which means half your library is invisible to any search tool. EnergyFile runs OCR on ingest so a scanned report from five years ago is just as queryable as something written yesterday.
The same gap between a list and a real search shows up in research data, which we covered in why a peptide list isn't a search tool.
Frequently asked questions
Why can't I trust an AI chatbot's answer about my documents?
Because there is no way to check where the answer came from without rereading the whole document, which defeats the point of asking. A general chatbot gives a confident, well-written answer either way. Without a citation it is just a guess with good grammar, and you may only find the error three decisions later.
How does semantic search find passages that use different words?
Documents are indexed with vector embeddings, so each chunk of a PDF, spreadsheet or presentation is encoded by meaning. A question can then match a passage that shares none of its words. Asking about exposure on the northern pipeline can find a paragraph that never uses the word exposure.
Can AI document search read scanned PDFs?
Only if OCR runs when documents are ingested. Many research libraries are half searchable text and half scanned pages that were never OCR'd, so half the library is invisible to any search tool. With OCR on ingest, a scanned report from five years ago is as queryable as one written yesterday.
