Skip to content

FLAVOR_PRESETS_GUIDE.md

Latest commit

 

History

History
95 lines (75 loc) · 2.27 KB

FLAVOR_PRESETS_GUIDE.md

File metadata and controls

95 lines (75 loc) · 2.27 KB

Flavor Presets Guide

Flavor presets are pre-configured templates for service flavors. Apply a preset to quickly configure processing mode and parameters.

Built-in Presets

Preset Mode Description
single-pass single_pass Process entire document in one LLM call. Auto-fallback to iterative if too long.
iterative iterative Process in batches with rolling context. For long documents.

single-pass

{
  "temperature": 0.2,
  "top_p": 0.8,
  "processing_mode": "single_pass",
  "auto_fallback_to_iterative": true
}

iterative

{
  "temperature": 0.2,
  "top_p": 0.7,
  "processing_mode": "iterative",
  "reduce_summary": true,
  "consolidate_summary": true,
  "create_new_turn_after": 150,
  "summary_turns": 3,
  "max_new_turns": 12
}

Using Presets

Web UI

  1. Navigate to Services > select a service > Add Flavor
  2. Select a preset from the dropdown
  3. Customize settings as needed
  4. Save

API

# List presets
curl http://localhost:8000/api/v1/flavor-presets

# Apply preset to create a flavor
curl -X POST "http://localhost:8000/api/v1/flavor-presets/{preset_id}/apply" \
  -H "Content-Type: application/json" \
  -d '{"service_id": "uuid", "model_id": "uuid", "flavor_name": "my-flavor"}'

Creating Custom Presets

Add a directory to seeds/presets/ with a manifest.json:

{
  "name": "my-preset",
  "description": {
    "en": "English description",
    "fr": "Description en francais"
  },
  "version": "1.0.0",
  "is_system": false,
  "config": {
    "temperature": 0.3,
    "processing_mode": "single_pass"
  }
}

Then run python -m app.seeds.base_seed --only presets to load.

Configuration Fields

Field Description
temperature Creativity (0.0-2.0)
top_p Nucleus sampling (0.0-1.0)
processing_mode single_pass or iterative
output_type text or json
auto_fallback_to_iterative Auto-switch if content exceeds context
create_new_turn_after Token threshold for splitting turns
summary_turns Context turns to keep
max_new_turns Turns per batch
reduce_summary Enable final consolidation step

See Swagger for full API reference.