| name | clickhouse-llm-chat |
| description | Execute LLM inference directly in ClickHouse using OpenAI, Claude, or Perplexity via Python UDFs. Use for text analysis, classification, summarization, or Q&A systems at scale. |
ClickHouse LLM Chat Skill
Purpose
Perform LLM inference operations (text generation, classification, analysis) directly within ClickHouse SQL queries using multiple LLM providers with intelligent retry logic and caching.
When to Use
- Perform sentiment analysis on customer reviews
- Classify support tickets or documents
- Generate summaries of long texts
- Extract entities or key information
- Build Q&A systems with RAG
- Analyze text at scale (millions of rows)
Supported Providers
- OpenAI (
openai_chat) - GPT-4, GPT-3.5-turbo
- Anthropic (
claude_chat) - Claude 3.5 Sonnet
- Perplexity (
perplexity_chat) - Sonar with web search
Quick Start
SELECT openai_chat('Analyze sentiment: ' || review_text) AS sentiment
FROM customer_reviews
LIMIT 100;
SELECT
ticket_id,
claude_chat('Classify as: technical, billing, general:\n' || message) AS category
FROM support_tickets;
Provider Selection Guide
| Provider | Speed | Cost | Quality | Use Case |
|---|
| GPT-3.5 | Fast | $ | Good | High-volume, simple |
| GPT-4 | Medium | $$$ | Excellent | Complex reasoning |
| Claude 3.5 | Fast | $$ | Excellent | Analysis, structured |
| Perplexity | Fast | $ | Good | Web-grounded info |
Best Practices
- Cache everything: LLM responses rarely change
- Batch processing: 10-100x more efficient
- Right model for task: Don't use GPT-4 for simple classification
- Structured prompts: JSON output for easier parsing
- Monitor costs: Track usage per provider
References
- See
../../src/skills/llm_skill.py
- See
../../src/tools/prompt_templates.py
- See
../../examples/sentiment_analysis/
