Search that answers the question, and shows its working
Readers do not arrive with keywords. They arrive with a problem, in their own words, often misspelled. AI search meets them there - and links to the article it used.
- Citations on every answer
- Permission-aware
- 60+ languages
How does AI search in TheDocs work?
When a reader asks a question, TheDocs retrieves the most relevant passages from the articles that reader is permitted to see, then generates a short answer from those passages and displays the source articles alongside it. Nothing outside your published content is used. If retrieval returns nothing relevant, the reader is told so and offered a handoff to support rather than given a guess.
- Retrieval is scoped to the reader's permissions and chosen version.
- Every answer links to the articles and sections it came from.
- Unanswerable questions are logged as content gaps, not hidden.
- Keyword results appear alongside the answer, never replaced by it.
- Answers stream, so the first words appear in well under a second.
- AI can be disabled per project if policy requires it.
What happens between the question and the answer
-
01
The question arrives
In the portal search box, the in-app widget, or through the API. Typos, jargon and half-sentences are normal input, not edge cases.
-
02
Permissions are applied first
The candidate set is narrowed to articles this reader may see, in the version and language they are browsing - before any retrieval happens.
-
03
Relevant passages are retrieved
Semantic and keyword retrieval run together, so both a paraphrased question and an exact error code find the right passage.
-
04
The answer is written and cited
A short answer is generated strictly from those passages, streamed to the reader, with links to each source article and section.
Why grounding matters more than model choice
A wrong answer costs more than no answer
In a help centre, a confident invention is worse than a blank result: the reader acts on it, it fails, and now you have a ticket and a trust problem.
- Generation is constrained to retrieved passages from your own content.
- Every claim in an answer is traceable to an article and section.
- Low-confidence retrieval produces an explicit 'not found' plus a handoff, configurable per project.
- Answers can be capped in length so the reader is pushed to the source for anything nuanced.
You can review recent answers and their sources in the admin console, and flag any that read wrongly - flagged answers surface the underlying article for editing.
Your search log is your content backlog
The most valuable output of AI search is not the answers. It is the list of questions your documentation cannot answer.
- Every query is recorded with whether it could be grounded.
- Ungrounded questions are clustered by topic, so twenty phrasings of the same gap appear as one item.
- Ranked by volume, so you fix the expensive gaps first.
- Create a draft article straight from a gap, pre-filled with the question as the title.
The same engine, everywhere readers are
A reader who has already opened a support form should not have to go looking for the help centre.
- Portal search box - the default reader experience.
- In-app widget - a search overlay you embed in your product.
- Support form - suggested answers as the customer types their subject line.
- API - query the same retrieval and answer pipeline from your own systems.
Related features
Questions people ask before they start
Ask your own documentation a hard question.
Start free for 14 days. No credit card, no setup fee, and your content is yours to export at any time.
Questions first? Email sales@thedocs.in or call +91 8585953085.