Ask your responses a question
Type what you want to know, the way you would say it to a colleague. Choice questions are answered exactly, by value. Free text is ranked by meaning, so "late delivery" finds "the courier came three days after the promised date" and "डिलीवरी में देरी" alike.
Six searches, and what comes back
From the evaluation on 200 sample responses about a delivery business.
Who can we contact?
Answered by value
Picks the question "Can we contact you?" and selects Yes. Shows everyone who said yes, with a count for every value so you can flip to No with one click.
unhappy customers on the business plan
Answered by value
Recognises the satisfaction question, selects Unhappy and Very unhappy, and lets you narrow by the plan question with a click.
late delivery
Ranked by meaning
Returns every complaint about delivery, in English, Hindi, and Hinglish, and nothing else. In the evaluation, all 31 relevant responses and no others.
डिलीवरी में देरी
Ranked by meaning
The same 31 responses. The query language does not have to match the answer language.
export data to Excel
Ranked by meaning
Finds the feature requests about spreadsheets, CSV, and Google Sheets, whether or not they said Excel.
recipe for pancakes
Nothing, correctly
Nothing. Unrelated questions return no results instead of the nearest noise.
Exact where it can be, meaning where it must be
- Choice questions are answered by value
- Embeddings are poor at yes and no: "No" sits almost as close to a query as "Yes". So when your question is about a choice field, the search picks the field and the values, and the results come from the reporting tables, exactly. You see every value with its count and can change the pick.
- Free text is ranked by meaning
- Each answer is embedded once when it arrives. A search embeds your question, finds the closest answers, and stops at the gap where related answers end. Two phrasings of the same complaint land next to each other; unrelated answers stay out.
- Measured, not promised
- The thresholds were set on a 200-response sample with labelled topics in English, Hindi, and Hinglish. 19 of 20 evaluation queries pass, with 100% precision and recall on delivery, crashes, exports, WhatsApp, and city queries. The one that fails is a vague phrasing; the evaluation script ships in the repository.
- Results are responses, not snippets
- Every result is a full response with the matching answers highlighted, so you can open it, export the set, or filter further.
Where the text goes
Search is optional per deployment. When it is on, answer text is sent to Voyage AI to be turned into vectors, processed in the United States under zero-day retention, and the vectors are stored in your own database. Your consent notice should say so, and the DPDP page explains how. Turn search off and nothing leaves the database.
Questions people ask
Which languages does search understand?
English, Hindi, and mixed Hinglish were evaluated. The embedding model is multilingual, so other languages work, but they have not been measured yet.
Does search see file uploads or numbers?
Numbers, dates, and choices are searched by value through the reporting tables. Files are not indexed.
Can an AI client search responses over MCP?
Not yet. MCP tools cover forms and drafts only, by design. Search is available in the app and the API.
Try it on your own answers
Create a workspace, import a CSV of past responses, and ask it the question you could never answer with a filter.