Route a multilingual helpdesk with typed triage

01 · Role in the system

What Jev does here

Each student message triggers two parallel Jev calls. Triage sees only the query and returns four atomic judgments: language (Choice), intent (Choice with confidence), urgent (Noul) and wants_human (Noul). A second call pairs the query with each retrieved passage and returns one grounding Noul per passage. Code composes the answers deterministically: hand off to a human when one is wanted or confidence is low, answer from a template when out of scope, and call the generative LLM only when intent is confident and passages are grounded - the LLM writes, Jev decides whether it may write at all. The repo also carries an author-run benchmark of this pipeline on low-resource languages: accuracy per language, calibration as ECE with a reliability diagram, exploratory intent coverage, latency and cost.

02 · Control boundary

Where Jev sits

Jev as the decision layer of a RAG helpdesk: atomic judgments composed by code, triage never sees retrieved passages, confidence acts as a second axis on intent, and fee arithmetic and deadlines stay in code.

Code owns the loop, permissions, thresholds, validation, and side effects. Jev owns only the bounded judgments described above.

03 · Known limits

What this evidence does not prove

• Benchmark results are author-run and regenerate locally via benchmark.py; summary outputs are not committed to the repo.
• The demo runs as a local Gradio app; there is no hosted public demo.
• Only English, Uzbek and Russian are benchmarked; other low-resource languages are untested.

04 · Attribution

Public sources

This is a Community record: the project was published by a third-party community author.

• rajantripathi ↗Community · github · public · checked 2026-09-27