jeff-adapter-assistant-routing

Assistant reasoning check. Predicts whether an assistant's next response requires careful reasoning or can be answered directly. Use it to route simpler turns to a faster, cheaper response path.

A LoRA adapter for jeff-base v1.4, Jeff's small open decision model (a fine-tune of Qwen3.5-0.8B). You give it a situation and a question with named options; it returns a calibrated probability for every option from one forward pass, with no generated text to parse.

• Adapter ID: jeff.agents.assistant-routing, version 1.4.0; short names: assistant-routing
• Catalogue page: jeffhub.ai/catalogue/jeff.agents.assistant-routing
• Code, server and SDKs: github.com/firelex/jeff
• Base it needs: mstrasser/jeff-base at revision v1.4. The engine refuses an adapter trained on another base, so this adapter loads only on that revision.

When to use it

• Your assistant has a separate reasoning step and needs to decide when to use it.

When not to use it

• You need to choose a tool or specialist; this adapter only checks whether reasoning is needed.

What it decides

What a call sends:

Field	Required	What it holds
accounts (context)	yes	The user's accounts the assistant can act on, as one text (for example "Work account [work], Personal account [personal]").
date_and_time_zone (context)	yes	Today's date and time and the user's time zone, as one sentence.
conversation (context)	yes	The conversation so far as one text, including the user's messages and the tool calls and results.
turn (input, fields)	yes	What the main assistant must do next in this turn (for example "Reply to the user in words.").

jeff.agents.assistant-routing.needs_thinking (detect, noul)

Instructions: Does the main assistant need to think carefully before writing this turn, rather than answering straight away?

Labels:

• A yes/no question: the answer is true or false.

How it was measured

Base: Jeff v1.4; adapter trained 2026-10-25.

Test	Rows	Accuracy
held-out simulated conversations (generator)	771	86.6%

Accuracy is the share of test rows whose top answer was right. Kinds: public, a public benchmark; real, held-out real documents; generator, an in-house test from the same generator as the training data (it measures less than a public or real test does); mixed, several kinds; unknown, not recorded.

When it is confident (from the calibration set): the threshold is the confidence from which an answer counts as confident, precision the share of confident answers that were right, coverage the share of answers that were confident.

Function	Threshold	Precision	Coverage	Calibration rows
needs_thinking	0.882	95.2%	54.8%	341

Caveats:

• The label is a stand-in: whether a large LLM chose to think before answering, not a human judgement.

How to call it

Install the Python SDK and the engine (see the install guide; until the packages are on PyPI, install them from github.com/firelex/jeff).

The example below is "Thanks after a moved meeting". The user only says thank you after a finished task, so the reply needs no careful thinking and the answer is false.

On your machine (Jeff.local downloads this adapter and the base from Hugging Face the first time):

import jeff

client = jeff.Jeff.local(["assistant-routing"])

decide = client.function("jeff.agents.assistant-routing.needs_thinking", context={
    "accounts": "Work account [work], Personal account [personal]",
    "date_and_time_zone": "It is Tuesday 3 November 2026, 16:20, and the user is in London (UK time).",
    "conversation": "User: Can you move my 4 pm call with Priya to Thursday at the same time? Assistant: (calls calendar.move_event) Result: the call is now on Thursday 5 November at 16:00. Assistant: Done, your call with Priya is now on Thursday at 4 pm. User: Great, thanks!",
})

decision = decide("Reply to the user in words.")

print(decision.value, decision.confidence)

Through the Jeff API (an API key from jeffhub.ai):

import os

import jeff

with jeff.Jeff.remote("https://api.jeffhub.ai", token=os.environ["JEFF_API_KEY"], timeout_s=30) as client:
    decide = client.function("jeff.agents.assistant-routing.needs_thinking", context={
        "accounts": "Work account [work], Personal account [personal]",
        "date_and_time_zone": "It is Tuesday 3 November 2026, 16:20, and the user is in London (UK time).",
        "conversation": "User: Can you move my 4 pm call with Priya to Thursday at the same time? Assistant: (calls calendar.move_event) Result: the call is now on Thursday 5 November at 16:00. Assistant: Done, your call with Priya is now on Thursday at 4 pm. User: Great, thanks!",
    })

    decision = decide("Reply to the user in words.")

    print(decision.value, decision.confidence)

curl -s https://api.jeffhub.ai/v1/decide \
  -H "Authorization: Bearer $JEFF_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"calls": [{"function": "jeff.agents.assistant-routing.needs_thinking", "input": "Reply to the user in words.", "context": {"accounts": "Work account [work], Personal account [personal]", "date_and_time_zone": "It is Tuesday 3 November 2026, 16:20, and the user is in London (UK time).", "conversation": "User: Can you move my 4 pm call with Priya to Thursday at the same time? Assistant: (calls calendar.move_event) Result: the call is now on Thursday 5 November at 16:00. Assistant: Done, your call with Priya is now on Thursday at 4 pm. User: Great, thanks!"}}]}'

Recorded answer (2026-11-03): false with confidence 0.951.

To download the weights yourself:

hf download mstrasser/jeff-base --revision v1.4 --local-dir jeff-base
hf download mstrasser/jeff-adapter-assistant-routing --revision v1.4 --local-dir adapters/assistant-routing

Files

File	What it is
adapter_model.safetensors, adapter_config.json	The LoRA weights and their configuration
readout.safetensors	The adapter's readout over the answer codes
decision_config.json	Answer codes, temperature, prompt layout and the base it was trained on (the SHA-256 of the base's weights, which the engine checks)
manifest.yaml	The adapter's manifest: its functions, labels, fields and calibration, as the SDKs read it
test.jsonl	The held-out test set the numbers above come from
calibration.jsonl	The calibration set the thresholds come from
LICENSE	The Apache License 2.0, for the Qwen3.5-0.8B notice below

Licence

Adapter licence: Apache-2.0

Qwen3.5-0.8B notice: these weights were modified from Qwen3.5-0.8B by the Jeff project: jeff-base is a fine-tune of Qwen3.5-0.8B, and this adapter was trained on top of it. Qwen3.5-0.8B is Copyright 2026 Alibaba Cloud and licensed under the Apache License, Version 2.0; a copy of that licence is in LICENSE.

Jeff is an independent project, not affiliated with or endorsed by the makers of Qwen.