LLM-based sentiment classification for social mentions

[arberx]

Summary

Add optional LLM-based sentiment classification as a post-processing step for social mentions.

Details

After social mentions are ingested, optionally classify each mention's sentiment as positive, neutral, or negative using a cheap LLM (Haiku or GPT-4o-mini).

Implementation

• Reuse existing provider infrastructure (ProviderAdapter.generateText()) — no new LLM integration needed
• Process new mentions in batch after each social monitoring run
• Store result in social_mentions.sentiment column (already in schema)
• Configurable via socialConfig.sentimentAnalysis: true in canonry.yaml
• Include confidence threshold — only store classification above threshold

Prompt design

Simple zero-shot classification prompt:

Classify the sentiment of this social media post about [brand] as positive, neutral, or negative. Respond with only the sentiment label.

Post: [content_text]

Depends on

• Social contracts, DB migration, at least one adapter

Acceptance Criteria

• Sentiment classification runs as post-processing after social runs
• Uses cheapest available provider (Haiku > GPT-4o-mini)
• Configurable on/off via socialConfig.sentimentAnalysis
• Confidence threshold prevents low-quality classifications
• Tests verify classification and storage