Is the 6/6 routing accuracy a real result or a demo number?

It is a real measurement of a very small set. check_jev.py asks six Chinese questions — two ML concepts, a calculation, a weather question, a greeting and a model-selection question — and each returned the expected label at confidence 1.0. That is a sanity check on the integration, not a benchmark; the README makes no claim about other languages or harder phrasings.

What stops the model from inventing a fifth route?

The shape of the request. Routes are declared as the criteria map on a single choice question, so the candidate set is fixed by you and the model can only pick from it. That is the stated point of the design — a typed label plus per-candidate probabilities and a confidence value, with no generated text anywhere in the response.

What happens if the Jev API is down or unsure?

Routing degrades to local keyword rules: weather words to weather, a verb plus digits to calculator, greetings to chat, everything else to knowledge. Anything under 0.6 confidence takes the same path. That is deliberate — the README argues against falling back to a three-tool agent deciding for itself, which is the roughly one-in-three behaviour the design set out to remove.

Does moving routing to Jev make the answers better?

Not directly, and the README does not claim it does. Routing only decides which pipeline runs; the knowledge route still answers from retrieved chunks with a bare LLM plus a RAG template, so answer quality rides on the corpus you index with bge-m3 into ChromaDB. What routing fixes is silent failure: the 3b model used to answer some concept questions from memory about a third of the time instead of retrieving.

What stack and models does it run on?

LangChain v1 create_agent on LangGraph, Ollama qwen2.5:3b for chat and bge-m3 for 1024-dim embeddings, ChromaDB persisted to .chroma/, Jev for routing, and Flask 3 with SSE streaming.

What do the local keyword rules do when Jev is skipped?

They route by pattern: weather words go to weather, a verb plus digits goes to calculator, greetings go to chat, and everything else goes to knowledge; Jev is accepted only at confidence 0.6 or higher.

How many LLM calls does each route make?

Knowledge is one deterministic retrieval call, chat is one bare LLM call, and calculator and weather are single-tool agents at two LLM calls each, since every route carries exactly one tool and one prompt.

CommunityJev Agents & MCP

Jev Routed Agent

A LangChain agent where Jev routes intent as a typed choice over four fixed criteria, with confidence below 0.6 falling back to local keyword rules, and each route carrying exactly one tool and one prompt.

See highlights

Open github.com

Stars: 4
Language: Python
Published by: Community
Author: liuhongrui087-art
Last verified: Oct 2026

liuhongrui087-art/jev-routed-agent

User repository on GitHub

View on GitHub

Multi-step reasoning agent built on LangChain v1 + Jev + Flask + Ollama, with tool calling and local RAG Q&A.

Forks: 1
Last pushed: Updated Sep 2026

Repo stats from the GitHub API, cached Oct 2026.

About this project

English | 简体中文 A multi-step reasoning agent built with LangChain v1 + Jev + Flask + Ollama. User question → the routing layer determines the intent → the matching pipeline (knowledge-base retrieval / math / weather / small talk) → a token-by-token streaming answer backed by sources. Dependencies flow one way: web → services → core → config.py. No HTTP-related code is allowed inside core/.

Quoted from the project's README. Read the full README.

Tagscommunity jevbest catalog-import

Works forJev for Classification Jev for Routing Jev for Tool calling

Used inJev for Software & DevTools

Languages: Python

Highlights

Six Chinese questions route 6/6 at confidence 1.0, against roughly one in three for the 3b model choosing its own tool.
Four routes — knowledge, calculator, weather, chat — each with exactly one tool and one prompt, so the model fills in arguments instead of choosing.
No parsing and no invention: the criteria map is the candidate set, so the model cannot return a fifth option, and probabilities come back per candidate.
Jev is optional. Low confidence or an API failure drops to local keyword rules rather than to a multi-tool agent deciding for itself.

Watch out

The headline routing result is six Chinese questions in check_jev.py, all answered at confidence 1.0, with recorded latencies between roughly one and two seconds — a small sample in one language, and every routing call is an extra network round trip in front of a local 3b model that is otherwise fast. Only Choice is used even though the decision model also offers Noul and Score, so the rest of the API is untouched. The knowledge route does not improve the corpus: it retrieves from a ChromaDB index you build yourself with bge-m3 embeddings persisted to .chroma/, and the determinism argument in the README is about routing, not about whether the retrieved chunks answer the question. The calculator and weather routes are single-tool agents that still spend two model calls, and the get_weather tool's data source is not described. If Jev is unavailable or returns confidence below 0.6 the system falls back to local keyword rules, which are readable but crude, so behaviour changes shape without a loud signal. Copy reviewed against the repository README on 2026-10-04, not from running it.

Questions about this entry

Short answers drawn from the linked project or page.

More Jev resources like this

Tiermem

Tiered agent memory keeping compact summaries with source links over searchable raw history, where a router decides per query whether notes suffice or raw pages must be recovered.