Question 1

What is code_2, the triage pipeline?

Accepted Answer

A small support-ticket triage pipeline in Python. Every incoming ticket is first judged by Jev, TypeSafe's System One model, which returns numbers and labels in a single call; plain Python if statements then decide whether the ticket is worth a second pass by a general LLM that writes a reply.

Question 2

How is Jev different from the LLM here?

Accepted Answer

As the README puts it: Jev judges, classifies and scores, returning typed values — probabilities, labels, scores — while the LLM reasons, writes and explains in free text. Jev runs on every ticket and produces a number you can compare with a threshold; the LLM is only called for tickets Jev flags.

Question 3

Which Jev primitives does the pipeline use?

Accepted Answer

Noul and Choice, with Score listed as not used yet. Noul answers how likely something is true — is_urgent, is_angry, needs_a_human — accessed as r.nouls[name].noul, a 0-to-1 float; Choice picks among labels — intent among refund, technical_help, spam and other — accessed as r.choices[name].choice, a string.

Question 4

When is the LLM actually called?

Accepted Answer

Only when Jev's scores cross the thresholds the code defines. Because the outputs are typed, the gate is a plain if statement rather than parsing generated text; when it escalates, handle_reply uses an OpenAI or Anthropic client to write the reply.

Question 5

How many API calls does one ticket cost?

Accepted Answer

One Jev call: all questions go in a single system_one request, so four judgments cost one round trip. The second pass is an extra LLM call only for escalated tickets, which is the point of the pattern — Jev decides cheaply, the LLM is reserved for the tickets that deserve it.

Question 6

What do I need to run it?

Accepted Answer

Python and the keys config.py expects in .env: TYPESAFE_API_KEY for Jev, plus OPENAI_API_KEY or ANTHROPIC_API_KEY for the reply step. The digest's captured README was truncated before the Running it section, so the exact commands are in the repository's own setup instructions.

Question 7

Is it production-ready?

Accepted Answer

No — it is a teaching pipeline: ten sample tickets in messages.py, illustrative scores in the README and no benchmarks. Treat it as a readable reference for the decide-with-Jev, write-with-an-LLM pattern rather than a production ticketing system.

code_2 — Fast decisions with Jev

Highlights

Questions about this entry

Reactions & coverage

From the Jev guides

From the Jev AI model community