AI Models¶

An AgentModel (in llms/) names a model so workflows can refer to it. There are two types: llm (the default) for llm and loop agents, judges and spec analysis, and decision for the decide engine. tuvl init --preset gemini|openai|anthropic|ollama writes llms/default.yaml and llms/analysis.yaml for you.

LLM models¶

kind: AgentModel
version: v1
metadata:
  name: default
spec:
  model: gemini/gemini-3.1-flash-lite     # a LiteLLM model id
  api_key: ${GEMINI_API_KEY}             # optional; read from the environment
  temperature: 0.2
  max_tokens: 2048
  timeout: 60

Decision models¶

kind: AgentModel
metadata: { name: triage-classifier }
spec:
  type: decision
  provider: litellm            # laya | jev | litellm
  model: gemini/gemini-3.1-flash-lite

A decision model picks one value of a decide agent's enum output with a confidence; it never writes free text.

Laya (self-hosted). tuvl init --decision-model laya writes this model, LAYA_API_* entries in .env, and a compose.yaml that runs the laya-serve container on localhost:8000 (docker compose up -d laya). tuvl ship adds Laya to the Helm chart (Shipping to production).

The tuvl/laya-serve image comes in two variants:

tuvl init scaffolds model: english with the -en image, and tuvl ship picks the image (and the Helm memory requests) that matches the model. An -en image has no multilingual weights and runs offline, so non-English text sent to it fails with HTTP 500 inference failed — set model: auto and use the full image when you need other languages.

kind: AgentModel
metadata: { name: laya }
spec:
  type: decision
  provider: laya
  model: english               # english | auto | multilingual | typed-decisions
  api_base: ${LAYA_API_BASE:http://localhost:8000}
  api_key: ${LAYA_API_KEY:}    # optional; an ${ENV_VAR} reference, never a literal key

Jev (hosted). provider: jev with its api_base, and model pinned to a version (jev-1.13.0).

OpenAI-compatible endpoints (a LiteLLM proxy, your own classifier) use provider: litellm with model: openai/<name>, api_base and an ${ENV_VAR} api_key. The endpoint must return {"decision": …, "confidence": 0..1} (or token logprobs); an answer with no confidence is an error, never treated as certain.

See Decide.

Where models are used¶

Model ids are pinned in tuvl.lock by tuvl lock. You can also manage models on the Insight AI Models page.