models.py

Go to the documentation of this file.

#!/usr/bin/env python3
# -*- coding: utf-8 -*-
# providers/ollama/models.py
 
"""
Ollama provider model registry.
 
Purpose
-------
:data:`MODELS` is the curated registry of well-known local models with
their capabilities. Live models discovered via ``/api/tags`` that are
not in this registry get :meth:`OllamaProvider.default_caps` instead
(permissive local defaults).
 
:data:`TIER_MAP` maps semantic tiers to model ids used by
:meth:`OllamaProvider.tier`.
 
Responsibilities
----------------
- :data:`MODELS` — ``dict[str, ModelCaps]`` for frequently-used local
  models across the llama, qwen, deepseek-r1, gemma, phi, and mistral
  families.
- :data:`TIER_MAP` — semantic-tier lookup.
 
Diagnostics
-----------
Domain: OC_CHAT.CLIENT
Levels:
    L2 — lifecycle events
    L3 — semantic details
    L4 — deep tracing
 
Contracts
---------
- All ``TIER_MAP`` values must be valid keys in ``MODELS``.
- Context windows reflect the *configured* Ollama defaults; actual
  context is capped by available VRAM and Ollama's ``num_ctx`` setting.
 
Dependencies
------------
- option_c_agent_contracts (ModelCaps)
- oc_diagnostics (structured debug logging via _dbg)
"""
from __future__ import annotations
 
from oc_diagnostics import _dbg
 
from option_c_agent_contracts import ModelCaps
 
_dbg("OC_CHAT.CLIENT", "module", "loaded", 4)
 
MODELS: dict[str, ModelCaps] = {
    # llama family
    "llama3.2":         ModelCaps(supports_thinking=False, supports_reasoning=False, supports_tools=True,  max_context_tokens=131_072),
    "llama3.2:1b":      ModelCaps(supports_thinking=False, supports_reasoning=False, supports_tools=True,  max_context_tokens=131_072),
    "llama3.2:3b":      ModelCaps(supports_thinking=False, supports_reasoning=False, supports_tools=True,  max_context_tokens=131_072),
    "llama3.1":         ModelCaps(supports_thinking=False, supports_reasoning=False, supports_tools=True,  max_context_tokens=131_072),
    "llama3.1:8b":      ModelCaps(supports_thinking=False, supports_reasoning=False, supports_tools=True,  max_context_tokens=131_072),
    "llama3.1:70b":     ModelCaps(supports_thinking=False, supports_reasoning=False, supports_tools=True,  max_context_tokens=131_072),
    # qwen family
    "qwen2.5":          ModelCaps(supports_thinking=False, supports_reasoning=False, supports_tools=True,  max_context_tokens=32_768),
    "qwen2.5:7b":       ModelCaps(supports_thinking=False, supports_reasoning=False, supports_tools=True,  max_context_tokens=32_768),
    "qwen2.5:14b":      ModelCaps(supports_thinking=False, supports_reasoning=False, supports_tools=True,  max_context_tokens=32_768),
    "qwen2.5:32b":      ModelCaps(supports_thinking=False, supports_reasoning=False, supports_tools=True,  max_context_tokens=32_768),
    "qwen2.5-coder:7b": ModelCaps(supports_thinking=False, supports_reasoning=False, supports_tools=True,  max_context_tokens=32_768),
    # deepseek-r1 local (reasoning, no tools in ollama)
    "deepseek-r1":      ModelCaps(supports_thinking=False, supports_reasoning=True,  supports_tools=False, max_context_tokens=65_536),
    "deepseek-r1:1.5b": ModelCaps(supports_thinking=False, supports_reasoning=True,  supports_tools=False, max_context_tokens=65_536),
    "deepseek-r1:7b":   ModelCaps(supports_thinking=False, supports_reasoning=True,  supports_tools=False, max_context_tokens=65_536),
    "deepseek-r1:8b":   ModelCaps(supports_thinking=False, supports_reasoning=True,  supports_tools=False, max_context_tokens=65_536),
    "deepseek-r1:14b":  ModelCaps(supports_thinking=False, supports_reasoning=True,  supports_tools=False, max_context_tokens=65_536),
    # gemma family
    "gemma3":           ModelCaps(supports_thinking=False, supports_reasoning=False, supports_tools=False, max_context_tokens=8_192),
    "gemma3:4b":        ModelCaps(supports_thinking=False, supports_reasoning=False, supports_tools=False, max_context_tokens=8_192),
    "gemma3:9b":        ModelCaps(supports_thinking=False, supports_reasoning=False, supports_tools=False, max_context_tokens=8_192),
    "gemma3:27b":       ModelCaps(supports_thinking=False, supports_reasoning=False, supports_tools=False, max_context_tokens=8_192),
    # phi family
    "phi4":             ModelCaps(supports_thinking=False, supports_reasoning=False, supports_tools=True,  max_context_tokens=16_384),
    "phi4:14b":         ModelCaps(supports_thinking=False, supports_reasoning=False, supports_tools=True,  max_context_tokens=16_384),
    # mistral family
    "mistral":          ModelCaps(supports_thinking=False, supports_reasoning=False, supports_tools=True,  max_context_tokens=32_768),
    "mistral:7b":       ModelCaps(supports_thinking=False, supports_reasoning=False, supports_tools=True,  max_context_tokens=32_768),
    "codestral":        ModelCaps(supports_thinking=False, supports_reasoning=False, supports_tools=True,  max_context_tokens=32_768),
    "codestral:22b":    ModelCaps(supports_thinking=False, supports_reasoning=False, supports_tools=True,  max_context_tokens=32_768),
    # nomic embed (not a chat model — listed for /models display)
    "nomic-embed-text": ModelCaps(supports_thinking=False, supports_reasoning=False, supports_tools=False, max_context_tokens=8_192),
}
 
TIER_MAP: dict[str, str] = {
    "default":   "llama3.2",
    "cheap":     "llama3.2",
    "reasoning": "deepseek-r1:8b",
    "summary":   "llama3.2",
}
 
__all__ = ["MODELS", "TIER_MAP"]