Search by meaning with Jev

Can Jev search by meaning instead of keywords?

Yes, within a bounded list. TypeSafe’s line-by-line search cookbook tags every line of a document with an id, sends the tagged text as state, and asks a Choice question: which line answers this query? The answer carries a probability for every line, so your code can rank them. The same shape fits any short list of text rows: the subjects in an inbox, the rows of a sheet, the clauses of a policy.

Open the editable line-search template. It holds five fictional shop-policy lines and a query that shares almost no words with the line that answers it. Loading it creates a draft; a check starts only when you request it.

Why the template has a “no line” option

Choice probabilities always add up to one. Some line therefore ranks first even when nothing in the document answers the query. The official cookbook handles this with a second, yes/no question asked in the same request: does any line address the query at all? Your code reads that answer before trusting the ranking.

The Playground asks one question at a time, so this template adds a sixth option, “No line answers it”, to teach the same habit. In an integration, prefer the cookbook’s two-question form.

Change one thing and compare

1. Run the template and note the probabilities for L2 and L3. Both mention returns, but only one covers an opened box.
2. Change the query to “How long does delivery take?” and run it again.
3. Ask something the lines cannot answer, such as “Do you ship to Canada?”, and see whether the last option wins.

No expected answer is prefilled. Three runs on five invented lines say nothing about accuracy on your documents.

Limits to plan around

• List size. The cookbook notes that a Choice question accepts up to 255 options, so one request covers up to 255 lines; longer documents need a first pass that narrows the candidates. This site’s editor allows 2–6 options, which is a website limit.
• Thresholds. The cookbook’s cut-offs for “answered”, “partial” and “absent” separate its own examples and are meant to be tuned on your documents.
• Retrieval still matters. TypeSafe’s re-ranking cookbook first builds a keyword shortlist and only then asks one question per query–passage pair. Jev judged the candidates it was given; it did not search the corpus.
• Facts stay in the text. A ranking points to a line. Show the line itself, not a paraphrase, so a reader can verify it.

The Projects gallery lists community search and re-ranking experiments; each is a single author-reported comparison rather than a benchmark. For mixed signals between close lines, read probability versus confidence.