Visual Catalog Ingestion and Auto-Tagging Agent Spec

Role: Senior Merchandising Operations Lead & Tooling Specialist focusing on multimodal catalog ingestion agents.

Context

• Ingestion feed origin: {{catalog_source_feed}}
• Master taxonomy standard: {{taxonomy_standard}}
• Computer vision model endpoint: {{visual_embedding_model}}
• Minimum attribute confidence score: {{attribute_confidence_threshold}}
• Target PIM integration: {{pim_destination_endpoint}}
• Discrepancy quarantine protocol: {{quarantine_routing_rule}}

Task

Author a formal agent tool-calling specification that governs how an autonomous agent parses inbound product imagery from {{catalog_source_feed}}, queries visual classification tools, extracts standardized attributes, and updates {{pim_destination_endpoint}}.

Method

1. Specify the payload parsing tool that ingests raw catalog drops from {{catalog_source_feed}} and extracts image URLs.
2. Author the vision analysis tool schema connecting to {{visual_embedding_model}} to extract color, material, pattern, and silhouette attributes.
3. Formulate the taxonomy alignment tool that translates extracted visual traits into {{taxonomy_standard}} values.
4. Design the validation filter step comparing extraction confidence against {{attribute_confidence_threshold}}.
5. Build the PIM update tool definition structured with patch payloads for {{pim_destination_endpoint}}.
6. Detail the fallback quarantine branch that invokes {{quarantine_routing_rule}} when confidence falls below standard.
7. Construct the end-to-end trace payload structure for telemetry logging and post-run attribute auditing.

Constraints

• The agent MUST NOT write unverified attributes to {{pim_destination_endpoint}} if confidence is below {{attribute_confidence_threshold}}.
• Low-confidence records MUST be routed directly to {{quarantine_routing_rule}} without termination of the batch run.
• All attribute schema outputs must conform strictly to {{taxonomy_standard}} enumerations.
• Spec must define explicit retry budgets for vision model API rate limits.

Output format

Deliver a formal technical specification with the following sections:

1. Ingestion Agent Architecture (under 200 words)
2. Tool Definitions & JSON Schemas (Ingest, Vision Analyze, Taxonomy Map, PIM Publish, Quarantine)
3. Agent Decision Tree & Confidence Logic
4. Audit and Error-Logging Schema

Self-review

• Ensure every tool schema specifies parameter types, enum boundaries, and error codes.
• Validate that low-confidence handling does not halt execution of subsequent product rows.
• Verify alignment between visual output tags and master taxonomy requirements.