Submission for the UP Hackathon, 11–12 September 2026, built on Apache Texera.
Texera ships 166 operators across 26 groups. The palette's search box compares what you type against the operator's name and nothing else, so a query phrased as an intent finds nothing — even when the operator's own description says exactly that. All 166 descriptions exist. None is ever read.
Put plainly: the current search only helps someone who already knows the answer.
This work adds search by meaning to that same box — no second mode and no switch to find. Both rankers answer every query: the keyword search keeps its places, so typing a name behaves exactly as before, and the ranked results fill what is left, which is where a phrase lands since it matches no name.
It is computed entirely in the browser: no backend, no API key, no per-query cost.
Across 40 natural-language queries, in English and in Spanish, the existing search returned zero results every time. Not a language problem — fuse.js matches the whole query against the operator name, so only the name works.
Queries that do name the operator still rank it first, 10 out of 10: the new ranking gives up nothing the old one could do.
Method, results and their limits: KADI/evaluacion.md.
Read KADI/metodo.md before quoting any accuracy figure — it
explains which numbers are trustworthy and which are not yet.
Vectors for all 166 operators are generated offline and committed to the repository. At runtime only the user's query is embedded, by a 23 MB model running in WebAssembly inside the browser, cached after first use.
Design decisions and what we tried that failed:
KADI/README.md.
The palette explains an operator until you drop it. After that the canvas shows a name and nothing else, and a renamed operator shows not even that — so reading someone else's workflow means clicking every box in turn.
Hovering an operator now gives its group and description together with its neighbours in this workflow:
Count per department
Aggregate
Calculate different types of aggregation values
──────────────────────────────────────────────
← from Drop unknown
→ to Busiest first, Share of tickets
The wiring is the half that cannot be read off the operator alone, and it comes from the workflow graph rather than a model: it cannot be wrong, and it updates as you connect and disconnect — including while the card is open, which is when it matters, since drawing a link never takes the cursor off the operator.
It stays quiet while the Performance overlay is on, which already owns the hover, and still renders for an operator whose type is missing from the metadata.
KADI/workflows/ holds a demo workflow. Import it from the
Workflows page with the upload button next to Create Workflow.
It sources its rows from a Text Input rather than a file, so it runs on a fresh
instance with no dataset uploaded. Count per department deliberately feeds two
branches, which is the operator worth hovering.
The palette's search already existed. We left it in place and added a second ranker beside it, so the two can be compared and either can be switched off.
1 · An offline index.
frontend/scripts/generate-operator-embeddings.mjs
reads the response of GET /api/resources/operator-metadata, builds one line of
text per operator from its name, group, description and any hand-written
phrasings, and embeds each one. The result is committed as
frontend/src/assets/operator-embeddings.json — 166 operators × 384 dimensions,
486 KB. This runs on a developer's machine, never in the app.
2 · A ranker in the browser.
semantic-operator-search.service.ts
loads that index, lazily imports Transformers.js on first use, embeds the user's
query and scores it against all 166 vectors by cosine similarity. Both sides are
normalised, so a dot product is the similarity. One shared warm-up promise means
a burst of keystrokes cannot start several model downloads.
3 · One box, two rankers.
operator-menu.component.ts
sends every keystroke to both. Keyword results render immediately, so the box
never waits on the model; ranked results merge in underneath once they arrive.
Queries carry a sequence number, so a slow answer for an earlier keystroke
cannot overwrite a newer one.
4 · The hover card. workflow-editor.component.ts builds it from the
operator's schema plus the graph's links, and subscribes to link, rename and
delete streams so an open card follows the graph instead of the cursor.
Nothing in the Texera backend was touched.
Type what you want to do, not what it is called. "remove repeated rows" rather than "Distinct". The box accepts both — an exact name still ranks first.
The first query in a fresh browser downloads the model (23 MB) and shows loading model…. After that it is cached and instant.
Each result carries three things: the operator's name, its relevance score, and its own description. The description is there so you can choose without opening anything.
Read the gap, not the number. The score is cosine similarity, not an
accuracy percentage. Distinct 0.71 above Limit 0.52 is a confident answer;
0.59 above 0.55 is a tie, and the descriptions are what settle it.
Clicking a result places the operator on the canvas, at the current viewport position — the same behaviour the old search had.
The original search is untouched. It runs on every query and its results come first; if the embedding model ever failed to load, they simply stand on their own.
Phrasings are how you improve it. KADI/operator-hints.json
maps an operator's display name to the wordings a user would reach for. Edit it,
re-run the generator, and the ranking changes. No model training involved.
A fork of Apache Texera (incubating), an open-source platform for human-AI collaborative data science through visual workflows.
Texera's own documentation lives in docs/; build and contribution
instructions in CONTRIBUTING.md.