e9e6fadb1d
- Fix ChatBubble to handle non-string content with String() wrapper - Fix API route to use generateText for non-streaming requests - Add @ai-sdk/openai-compatible for non-OpenAI providers (DeepSeek, etc.) - Use Chat Completions API instead of Responses API for compatible providers - Update ChatBubble tests and fix component exports to kebab-case - Remove stale PascalCase ChatBubble.tsx file
1.9 KiB
1.9 KiB
CSV Data File Standards
Purpose: When workflows need structured data that LLMs cannot generate.
When to Use CSV
Use CSV for data that is:
- Domain-specific and not in training data
- Too large for prompt context
- Needs structured lookup/reference
- Must be consistent across sessions
Don't use for:
- Web-searchable information
- Common programming syntax
- General knowledge
- Things LLMs can generate
CSV Structure
category,name,pattern,description
"collaboration","Think Aloud Protocol","user speaks thoughts → facilitator captures","Make thinking visible during work"
"creative","SCAMPER","substitute→combine→adapt→modify→put→eliminate→reverse","Systematic creative thinking"
Rules:
- Header row required, descriptive column names
- Consistent data types per column
- UTF-8 encoding
- All columns must be used in workflow
Common Use Cases
1. Method Registry
Advanced Elicitation uses CSV to select techniques dynamically:
category,name,pattern
collaboration,Think Aloud,user speaks thoughts → facilitator captures
advanced,Six Thinking Hats,view problem from 6 perspectives
2. Knowledge Base Index
Map keywords to document locations for surgical lookup:
keywords,document_path,section
"nutrition,macros",data/nutrition-reference.md,## Daily Targets
3. Configuration Lookup
Map scenarios to parameters:
scenario,required_steps,output_sections
"2D Platformer",step-01,step-03,step-07,movement,physics,collision
Best Practices
- Keep files small (<1MB if possible)
- No unused columns
- Document each CSV's purpose
- Validate data quality
- Use efficient encoding (codes vs full descriptions)
Validation Checklist
For each CSV file:
- Purpose is essential (can't be generated by LLM)
- All columns are used somewhere
- Properly formatted (consistent, UTF-8)
- Documented with examples