Per-Node Model Selection

This guide explains how to assign different LLM models to individual pipeline nodes, allowing fine-grained control over quality and speed.

Model Resolution Chain

Each pipeline node resolves its LLM model through a fallback chain. The first match wins:

Non-writer nodes (concept, culture, format_eval, cultural_eval, concept_eval, brand_eval):

1. Ad Unit override — pipeline_model_config[node_key] on the VideoAdUnit

2. PipelineSettings node default — per-node default in the singleton settings

3. PipelineSettings global default — app-level fallback for all nodes

4. Language primary model — the language’s configured LLM (ultimate fallback)

Writer node (different chain — prioritizes language over app settings):

1. Ad Unit override — pipeline_model_config["writer"]

2. Ad Unit LLM override — effective_llm_model FK on the ad unit

3. Language primary model — the language’s configured LLM

4. PipelineSettings global default — app-level fallback

The writer node defaults to the language’s primary model because it produces the target-language content and needs a model capable in that language.

Node Keys

Node Key	Pipeline Node	Role
concept	Concept Analyst	Extracts themes from origin script
culture	Cultural Researcher	Produces cultural brief for target market
writer	Script Writer	Adapts script for target audience
format_gate	Format Evaluator	Checks language compliance
culture_gate	Cultural Evaluator	Validates cultural sensitivity
concept_gate	Concept Evaluator	Verifies concept preservation
brand_gate	Brand Evaluator	Checks brand consistency

Setting Overrides at the Ad Unit Level

Each VideoAdUnit has a pipeline_model_config JSONField that maps node keys to LLM model primary keys:

{
  "concept": 2,
  "culture": 2,
  "writer": 5,
  "format_gate": 2,
  "culture_gate": 2,
  "concept_gate": 2,
  "brand_gate": 2
}

To set overrides:

1. Navigate to the Video Ad Unit in the admin

2. Edit the Pipeline Model Config JSON field

3. Use the primary key of the desired LLMModel for each node

4. Save — the override applies to this ad unit’s next pipeline run

Only include nodes you want to override. Omitted nodes fall through to PipelineSettings defaults.

Setting Defaults via PipelineSettings

PipelineSettings provides app-level defaults that apply to all adaptations unless overridden per-ad-unit:

1. Navigate to Core > Pipeline Settings

2. Select the LLM model for each node

3. Set the Global Default Model as the fallback for any unconfigured node

4. Save

See Pipeline Settings for the full PipelineSettings guide.

Practical Examples

Quality-focused configuration — use a larger model for evaluation nodes:

{
  "concept": 3,
  "culture": 3,
  "writer": 5,
  "format_gate": 3,
  "culture_gate": 3,
  "concept_gate": 3,
  "brand_gate": 3
}

Where model 3 is Qwen2.5-7B (higher quality) and model 5 is a language-specialized model for writing.

Speed-focused configuration — use a smaller model everywhere:

{
  "concept": 1,
  "culture": 1,
  "writer": 1
}

Where model 1 is Qwen2.5-3B (faster). Evaluation nodes fall through to PipelineSettings defaults.

Language-specialized writing — override only the writer for a market with specific language requirements:

{
  "writer": 7
}

Where model 7 is a model with strong capabilities in the target language. All other nodes use PipelineSettings defaults.

How Resolution Works

The resolve_pipeline_models() function in src/cw/lib/pipeline/state.py executes the resolution chain:

1. Loads the PipelineSettings singleton

2. Gets the ad unit’s pipeline_model_config overrides

3. For each node key, walks the fallback chain until a model is found

4. Returns a dictionary mapping node keys to {"model_id": str, "load_in_4bit": bool}

Only active models (is_active=True) are used. If a referenced model has been deactivated, the resolution falls through to the next level.