Answer strategies implement Retrieval-Augmented Generation (RAG) to answer user questions using documentation search and LLM inference.

Each answer strategy follows this general pattern:

1. Search: Use the active search strategy to find relevant documentation pages
2. Retrieve: Get full content from the top N search results
3. Format: Format the documentation as context for the LLM
4. Generate: Call the Groq LLM with the context and user question to generate an answer

The active strategy is selected in answer/index.ts by uncommenting the desired import.

llama-3.3-70b-default - Default strategy using Llama 3.3 70B

• Uses up to 5 documentation pages in context
• Truncates each page to ~2000 tokens to manage context window
• Temperature: 0.3 (relatively focused/deterministic)
• Model: llama-3.3-70b-versatile

Best for: General purpose Q&A with good balance of quality and speed

• Model: llama-3.3-70b-versatile
• Context: Up to 5 pages, 2000 tokens each
• Temperature: 0.3 (default)
• Performance: ~1-3s depending on search and LLM response time

Configuration options:

{
  limit: 10,              // Search results to consider
  minScore: 0,            // Minimum search score
  maxContextPages: 5,     // Pages to include in context
  temperature: 0.3,       // LLM temperature
  model: "llama-3.3-70b-versatile"  // Override model
}

To create a new answer strategy:

1. Create a new file in /answer/ (e.g., llama-3.1-8b-fast.ts)
2. Implement the AnswerStrategy interface:

import type { AnswerStrategy, AnswerResult, AnswerOptions } from "./types.ts";

export const answerStrategy: AnswerStrategy = {
  name: "your-strategy-name",
  description: "Brief description of your strategy",
  
  answer: async (query: string, options: AnswerOptions = {}): Promise<AnswerResult> => {
    // Your implementation here
    // 1. Search for relevant pages
    // 2. Get full content
    // 3. Format as context
    // 4. Call LLM
    // 5. Return AnswerResult
  },
};

3. Add the import to answer/index.ts and comment out other strategies
4. Test your strategy

Here are some ideas for additional strategies you might want to implement:

• llama-3.1-8b-fast: Faster responses with smaller model, fewer pages
• mixtral-8x7b-extended: More context pages with Mixtral's larger context window
• llama-3.3-70b-code-focused: Specialized prompts for code examples and API usage
• multi-step-reasoning: Break down complex questions into sub-questions
• citation-mode: Include specific citations/references in the answer
• conversational: Support follow-up questions with conversation history

GET /answer?q=How+do+I+use+the+Groq+API

GET /answer?q=What+models+are+available&maxContextPages=3&temperature=0.5

GET /answer/info

{
  "answer": "markdown formatted answer",
  "query": "user's original question",
  "searchResults": [
    {
      "path": "api/reference",
      "url": "https://...",
      "title": "API Reference",
      "score": 95.2
    }
  ],
  "contextUsed": 5,
  "totalTokens": 8500,
  "metadata": {
    "strategy": "llama-3.3-70b-default",
    "model": "llama-3.3-70b-versatile",
    "temperature": 0.3,
    "searchResultsCount": 10,
    "timings": {
      "search": 45.2,
      "contextPrep": 5.1,
      "llm": 1250.3,
      "total": 1300.6
    }
  }
}

• More pages = more comprehensive but slower and more expensive
• Fewer pages = faster but might miss relevant info
• Recommended: 3-5 pages for most questions

• Per page: 1000-3000 tokens is a good range
• Total context: Stay under 32k tokens for most models
• Llama 3.3 70B supports up to 128k context, but quality may degrade

• 0.0-0.3: Focused, deterministic answers (good for facts)
• 0.4-0.7: Balanced creativity and accuracy
• 0.8-1.0: More creative but less consistent (avoid for docs Q&A)

• Higher limit: More pages to choose from (10-20 recommended)
• Higher minScore: Only use highly relevant pages (50-70 threshold)
• Balance between precision (fewer, better pages) and recall (more pages, might miss relevant info)

1. Match model to task:

   • Simple factual questions: Smaller/faster models
   • Complex reasoning: Larger models (70B+)
   • Code generation: Models fine-tuned for code

2. System prompts matter:

   • Be specific about the documentation domain (Groq API)
   • Set expectations for answer format (markdown, citations, etc.)
   • Provide guidelines for handling missing information
   • Include URL formatting instructions (e.g., "remove .md extensions")

3. Manage context window:

   • Don't just dump all pages - be selective
   • Truncate long pages intelligently
   • Consider summarizing very long pages first

4. Error handling:

   • Gracefully handle search failures
   • Provide fallback responses when LLM fails
   • Include relevant metadata for debugging

5. Post-processing:

   • Use cleanupMarkdownLinks() to remove .md extensions from generated links
   • Add custom post-processing for your specific needs
   • Validate links and citations if needed

Answers automatically clean up markdown links to remove .md extensions:

// Before cleanup
[Compound](https://console.groq.com/docs/agentic-tooling/compound-beta.md)

// After cleanup  
[Compound](https://console.groq.com/docs/agentic-tooling/compound-beta)

This happens through:

1. System prompt: Instructs the LLM to avoid .md extensions
2. Post-processing: cleanupMarkdownLinks() function removes any remaining .md extensions

To disable cleanup (not recommended):

// In your custom strategy, skip the cleanup:
return {
  answer: llmResponse, // Don't call cleanupMarkdownLinks()
  // ... rest of result
};

Test your strategy with various question types:

• Factual: "What models does Groq support?"
• How-to: "How do I authenticate with the API?"
• Comparison: "What's the difference between model X and Y?"
• Code examples: "Show me an example of streaming responses"
• Complex: "How do I implement rate limiting with retries?"

Target performance for answer strategies:

• Search: 50-500ms (depends on active search strategy)
• Context prep: <50ms
• LLM call: 500-3000ms (depends on model and response length)
• Total: 1-4s for most queries

Strategies taking >5s should be optimized or marked as "detailed" mode.