6cab9e56fc
fix: model definition re-introduced into the schema but intelligently and only a summary is generated per tool. Required to ensure CLI calls and uses the correct model fix: removed `model` param from some tools where this wasn't needed fix: fixed adherence to `*_ALLOWED_MODELS` by advertising only the allowed models to the CLI fix: removed duplicates across providers when passing canonical names back to the CLI; the first enabled provider wins
70 lines
2.8 KiB
Markdown
70 lines
2.8 KiB
Markdown
# Model Capability Ranking
|
||
|
||
Auto mode needs a short, trustworthy list of models to suggest. The server
|
||
computes a capability rank for every model at runtime using a simple recipe:
|
||
|
||
1. Start with the human-supplied `intelligence_score` (1–20). This is the
|
||
anchor—multiply it by five to map onto the 0–100 scale the server uses.
|
||
2. Add a few light bonuses for hard capabilities:
|
||
- **Context window:** up to +5 (log-scale bonus when the model exceeds ~1K tokens).
|
||
- **Output budget:** +2 for ≥65K tokens, +1 for ≥32K.
|
||
- **Extended thinking:** +3 when the provider supports it.
|
||
- **Function calling / JSON / images:** +1 each when available.
|
||
- **Custom endpoints:** −1 to nudge cloud-hosted defaults ahead unless tuned.
|
||
3. Clamp the final score to 0–100 so downstream callers can rely on the range.
|
||
|
||
In code this looks like:
|
||
|
||
```python
|
||
base = clamp(intelligence_score, 1, 20) * 5
|
||
ctx_bonus = min(5, max(0, log10(context_window) - 3))
|
||
output_bonus = 2 if max_output_tokens >= 65_000 else 1 if >= 32_000 else 0
|
||
feature_bonus = (
|
||
(3 if supports_extended_thinking else 0)
|
||
+ (1 if supports_function_calling else 0)
|
||
+ (1 if supports_json_mode else 0)
|
||
+ (1 if supports_images else 0)
|
||
)
|
||
penalty = 1 if is_custom else 0
|
||
|
||
effective_rank = clamp(base + ctx_bonus + output_bonus + feature_bonus - penalty, 0, 100)
|
||
```
|
||
|
||
The bonuses are intentionally small—the human intelligence score does most
|
||
of the work so you can enforce organisational preferences easily.
|
||
|
||
## Picking an intelligence score
|
||
|
||
A straightforward rubric that mirrors typical provider tiers:
|
||
|
||
| Intelligence | Guidance |
|
||
|--------------|----------|
|
||
| 18–19 | Frontier reasoning models (Gemini 2.5 Pro, GPT‑5) |
|
||
| 15–17 | Strong general models with large context (O3 Pro, DeepSeek R1) |
|
||
| 12–14 | Balanced assistants (Claude Opus/Sonnet, Mistral Large) |
|
||
| 9–11 | Fast distillations (Gemini Flash, GPT-5 Mini, Mistral medium) |
|
||
| 6–8 | Local or efficiency-focused models (Llama 3 70B, Claude Haiku) |
|
||
| ≤5 | Experimental/lightweight models |
|
||
|
||
Record the reasoning for your scores so future updates stay consistent.
|
||
|
||
## How the rank is used
|
||
|
||
The ranked list is cached per provider and consumed by:
|
||
- Tool schemas (`model` parameter descriptions) when auto mode is active.
|
||
- The `listmodels` tool’s “top models” sections.
|
||
- Fallback messaging when a requested model is unavailable.
|
||
|
||
Because the rank is computed after restriction filters, only allowed models
|
||
appear in these summaries.
|
||
|
||
## Customising further
|
||
|
||
If you need a different weighting you can:
|
||
- Override `intelligence_score` in your provider or custom model config.
|
||
- Subclass the provider and override `get_effective_capability_rank()`.
|
||
- Post-process the rank via `get_capabilities_by_rank()` before surfacing it.
|
||
|
||
Most teams find that adjusting `intelligence_score` alone is enough to keep
|
||
auto mode honest without revisiting code.
|