Enrichment — AI Content Generation

Enrichment uses LLM to generate and enhance product content — descriptions, attributes, and wiki information. It complements Gardener (normalization) and Matcher (product linking) in the Commerce AI pipeline.

Description Generation

Generate professional, casual, or SEO-optimized product descriptions in any language.

Taxonomy Sync

Classify products into the site’s category hierarchy via Brain knowledge graph.

Wiki Enrichment

Research and populate brand pages, technology descriptions, and catalog entries.

Multi-language

Ukrainian-first (canonical) with LLM-powered translation to any target language.

Per-Task Model Configuration

Enrichment tasks use dedicated LLM models configured per-graph in contextunity.project.yaml:

# contextunity.project.yaml — multi-graph manifest
router:
  toolkits:
    - EnrichmentTools
  graph:
    enricher:
      template: "yaml:enricher"
      config:
        bilingual: true
      nodes:
        - name: "enrich"
          model: "openai/gpt-5-mini"
          model_secret_ref: "CU_ROUTER_GENERATION_MODEL_KEY"

    writer:
      template: "yaml:enricher"           # Same template, different prompts
      config:
        bilingual: true
        task: "description_generation"
      nodes:
        - name: "enrich"
          model: "openai/gpt-5-mini"
          model_secret_ref: "CU_ROUTER_GENERATION_MODEL_KEY"

The enricher and writer graphs both use the enricher YAML template but with different task configs. Enricher handles NER and attribute extraction; writer handles description generation.

Description Generation

Generate product descriptions via contextunity.router’s LLM pipeline:

# Via MCP tool
result = await enrich_product_description(
    product_id=42,
    language="uk",      # Ukrainian (canonical)
    style="professional"  # professional | casual | seo
)

Input context (automatically gathered):

Product title, categories, attributes
Existing description (if any)
Brand information

Output: 2-3 paragraphs highlighting key features, technical specs, and benefits.

Styles

Style	Use Case
`professional`	Product pages, catalogs — factual, structured
`casual`	Blog posts, social media — conversational tone
`seo`	Search optimization — keyword-rich, scannable

Language Support

The canonical language is Ukrainian (set in project config). Descriptions can be generated in any language — the system uses the project’s get_primary_language() to determine the default.

To generate translations after the canonical description exists:

Generate canonical description in Ukrainian
Use update_product() to save
Generate translated versions for other languages (en, de, etc.)
Save to Product.translations JSON field

Taxonomy Classification

Taxonomy sync pushes the site’s category tree to contextunity.brain for knowledge graph operations:

# Sync category taxonomy to Brain
python manage.py taxonomy_sync --domain category

# Sync color taxonomy
python manage.py taxonomy_sync --domain color

The taxonomy tree is also available via MCP:

tree = await get_taxonomy_tree(max_depth=3)
# Returns flat list of category nodes with depth and product counts

Enrichment Queue

Products needing enrichment are tracked by status:

Status	Meaning
`raw`	Newly harvested, no enrichment yet
`enriching`	Currently being processed
`enriched`	All enrichment tasks complete

Monitoring

# List products pending enrichment
queue = await list_enrichment_queue(dealer="gorgany", limit=25)

# Check enrichment status for a specific product
status = await get_enrichment_status(dealer_product_id=12345)
# Returns: normalized fields, enrichment metadata, trace_id

Wiki Enrichment

Wiki enrichment populates structured content for brands and technologies:

Brands

Company description and history
Product lines and specializations
Technology partnerships

Technologies

Technical description (Gore-Tex, Vibram, PrimaLoft, etc.)
Benefits and applications
Related products

Wiki content is managed through Wagtail CMS snippets (wiki.Brand, wiki.Technology) and enriched via the wiki_researcher node defined in the project manifest.

MCP Tools

Tool	Tags	Description
`enrich_product_description`	enrichment, mutate	Generate product description via LLM
`get_enrichment_status`	enrichment, query	Check enrichment status for a product
`list_enrichment_queue`	enrichment, query	List products pending enrichment
`get_taxonomy_tree`	enrichment, query	Get category tree with product counts

Architecture

Enrichment is fully manifest-driven. Commerce has zero dependency on contextunity-router:

Commerce (Django PIM)
   │
   ├── contextunity.project.yaml     ← declares enricher + writer graphs
   │
   ├── PIM Admin [Enrich] button
   │   └── calls export_product_for_enrichment (federated tool)
   │         │
   │         ▼
   │   Router loads enricher graph (yaml:enricher template)
   │         │
   │         ├── prepare (federated) → export product context via BiDi
   │         ├── enrich (platform: router_generate_content) → LLM generation
   │         ├── review (platform: router_review_content) → quality check
   │         └── update (federated) → write results back to Django ORM via BiDi
   │
   └── MCP tools for enrichment monitoring/triggering

Configuration

Commerce connects to Router using auto-discovery via Redis:

# Commerce .env
REDIS_URL=redis://localhost:6379          # Enables auto-discovery of contextunity.router
REDIS_SECRET_KEY=...                      # Required if Redis encryption is enabled

LLM models and secrets are configured per-graph in the contextunity.project.yaml manifest. See above.

File Locations

# Commerce — enrichment/writer federated tools
extensions/commerce/src/contextunity/commerce/
├── mcp/tools/
│   ├── enrichment.py                 # @federated_tool: export/update enriched products
│   └── gardener.py                   # @federated_tool: normalize, classify
├── pim/views/products/enrich.py      # PIM admin enrichment endpoints
├── harvester/management/commands/
│   └── taxonomy_sync.py              # Taxonomy sync to Brain
└── wiki/models.py                    # Brand, Technology snippets

# Router — enricher template (shared by enricher + writer)
services/router/src/contextunity/router/cortex/templates/
└── enricher.yaml                     # prepare → enrich → review pipeline