Could AI actually answer questions from this document?
Paste a policy, procedure or knowledge-base article. You'll get eight specific findings about whether that content can support a retrieval system — and the answer is often no, for reasons that have nothing to do with the model.
Nothing leaves your browser. No upload, no server, no analytics on the text, no storage. Read the source if you like — it's all in this one file.
Paste a document or pick a sample. The analysis appears here — eight checks, each with what it found and why it matters.
What actually determines whether a document is answerable
These are the properties that decide whether retrieval finds the right passage and whether a model can answer from it without inventing anything. None of them are about the model.
Length and structure
Documents get split into chunks before they're indexed. Too short and there's no context; too long with no headings and the split lands mid-thought, so retrieval returns half an answer.
Self-contained passages
"As mentioned above" and "this process" break when a chunk is retrieved alone. A passage that only makes sense in sequence will be quoted out of context.
Specifics over vagueness
"Promptly", "as appropriate", "at the manager's discretion" cannot be turned into an answer. Numbers, thresholds and named roles can.
Internal contradictions
Two different thresholds for the same thing in one document means the system will confidently pick one. This is the single most common cause of a wrong answer.
Time-bound claims
Undated references — "the new process", "from next quarter" — go stale invisibly. Without a date, nothing can tell whether the passage still applies.
Ownership and currency
A document with no owner and no review date can't be trusted or maintained. It's also the first thing an auditor asks about.
Question coverage
Does it actually answer the questions people ask? Documents often explain a policy thoroughly and never address the exceptions that generate every ticket.
What isn't checked
Whether the content is correct. No tool can tell you that, and any that claims to is guessing. Accuracy needs a person who knows the process.
What this tool can't tell you
It's a heuristic check on one document. Useful, and much narrower than a real assessment.
- Whether the content is true. The most important question and completely outside what any text analysis can determine. A beautifully structured document describing a process you abandoned in 2023 scores well here and is worse than useless.
- Whether it contradicts your other documents. Cross-document conflicts are the biggest real-world problem and need the whole corpus. This sees one file.
- Whether people can find it today. A perfect document nobody knows exists isn't answering anything.
- How your whole estate scores. One good document doesn't mean a good corpus. In practice the distribution matters far more than the best example.
- Whether retrieval would actually work. That needs embedding your real content and testing against real questions with known answers. This is a proxy for that, not a substitute.
The heuristics are deliberately conservative and will occasionally be wrong about your document — particularly with unusual formatting or non-English text. If a finding looks incorrect, it probably is; the tool doesn't get a vote about your business.
That's the normal result, and it's fixable
Most first documents score between 40 and 65. The useful move isn't rewriting everything — it's finding out which of your documents matter, which are contradicting each other, and whether the answers people need exist anywhere at all.
That's a two-week assessment across your real corpus, fixed fee, and it can conclude that you're not ready — which saves considerably more than it costs.
info@bshadmehr.meUseful in a first message
- Roughly how many documents you have, and where.
- The score this tool gave your typical one.
- The three questions people ask most often.
- Whether anyone owns documentation today.
- What you'd do with the time if it worked.