GitHub user gupta-sahil01 added a comment to the discussion: Ambient "operator
recommender" — predictive next-operator suggestions on the canvas
1) Offline analysis
Three small text files we write once and update as the operator catalogue
changes — no model training:
a short description of each operator (so the LLM knows what each one does)
a handful of example patterns ("after a Filter on text, suggest Sentiment") to
guide ranking
a small test set of workflows to check ranking quality in CI
Adding a new operator just means adding one line to the descriptions file. CI
catches it if you forget.
2) What's stored and where
Two places:
In the repo (versioned with code, reviewed in PRs): the three files above, plus
the LLM prompt templates. That's the entire "knowledge base."
In memory at runtime (auto-evicting cache): recent suggestion results, and
one-time analyses of any UDF source code we've seen. Caches a recent UDF
summary so we don't pay the LLM to re-read the same Python twice.
No database, no model registry, no separate service. Four files in the repo + a
cache.
GitHub link:
https://github.com/apache/texera/discussions/5240#discussioncomment-17154815
----
This is an automatically sent email for [email protected].
To unsubscribe, please send an email to: [email protected]
