This guide will help you quickly get started with answer strategies for RAG-based question answering.

Answer strategies implement Retrieval-Augmented Generation (RAG) - they search documentation, retrieve relevant pages, and use an LLM to generate answers to user questions.

# Run comprehensive test suite with timing breakdown
deno task answer

This will:

• Load test questions from testing/questions.ts
• Run each question through the active answer strategy
• Show detailed timing breakdown (search, context prep, LLM call)
• Display search results used for context
• Show generated answers
• Provide summary statistics

deno task serve

# Basic question
curl "http://localhost:8000/answer?q=How+do+I+use+the+Groq+API"

# With options
curl "http://localhost:8000/answer?q=What+models+are+available&maxContextPages=3"

# Get strategy info
curl "http://localhost:8000/answer/info"

# Run test suite via API
curl "http://localhost:8000/answer/test"

1. User asks a question: "How do I authenticate with the Groq API?"

2. Search (50-500ms): The active search strategy finds relevant docs

   → Search query: "authenticate groq api"
   → Returns: Top 10 matching pages ranked by relevance
   

3. Retrieve (fast, from cache): Get full content for top 5 pages

   → Loads content from latest.json (local cache)
   → Each page limited to ~2000 tokens
   

4. Format (<50ms): Package docs as LLM context

   → Formats pages with titles, URLs, and content
   → Adds system prompt with instructions
   → Creates user prompt with question
   

5. Generate (500-3000ms): LLM generates answer

   → Calls Groq API with Llama 3.3 70B
   → Returns markdown-formatted answer
   → Includes citations and examples from docs
   

{
  "answer": "To authenticate with the Groq API...",
  "query": "How do I authenticate with the Groq API?",
  "searchResults": [
    {
      "path": "api/authentication",
      "url": "https://console.groq.com/docs/api/authentication",
      "title": "Authentication",
      "score": 92.5
    }
  ],
  "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
    }
  }
}

• answer: The LLM-generated answer (markdown formatted)
• query: Your original question
• searchResults: Top pages used as context (with scores)
• contextUsed: Number of pages actually sent to the LLM
• totalTokens: Approximate tokens in the context
• metadata.timings: Performance breakdown

The active strategy is set in answer/index.ts:

// Default: Llama 3.3 70B with up to 5 doc pages
import { answerStrategy } from "./llama-3.3-70b-default.ts";

Currently available:

• ✅ llama-3.3-70b-default: Balanced quality and speed

Coming soon (you can implement these!):

• ⚡ llama-3.1-8b-fast: Faster responses with smaller model
• 📚 mixtral-8x7b-extended: More context with larger model
• 💻 llama-3.3-70b-code: Optimized for code examples
• 🔗 citation-mode: Includes specific citations

1. Comment out current strategy in answer/index.ts
2. Uncomment the strategy you want to use
3. Restart the server

// answer/index.ts

// Comment out default
// import { answerStrategy } from "./llama-3.3-70b-default.ts";

// Use fast strategy instead
import { answerStrategy } from "./llama-3.1-8b-fast.ts";

• q: The question to answer

  ?q=How+do+I+use+streaming

• limit: Max search results (default: 10)

  ?limit=20

• minScore: Minimum search score 0-100 (default: 0)

  ?minScore=50

• maxContextPages: Pages to send to LLM (default: 5)

  ?maxContextPages=3

• temperature: LLM temperature 0-1 (default: 0.3)

  ?temperature=0.5

• model: Override the model (optional)

  ?model=llama-3.1-8b-instant

# Fast mode: fewer pages, instant model
curl "http://localhost:8000/answer?q=What+is+Groq&maxContextPages=2&model=llama-3.1-8b-instant"

# High quality: more pages, higher threshold
curl "http://localhost:8000/answer?q=How+to+implement+retry+logic&limit=20&minScore=60&maxContextPages=7"

# Creative mode: higher temperature
curl "http://localhost:8000/answer?q=Ideas+for+using+Groq+API&temperature=0.8"

Good for: Quick facts, definitions, model names

curl "http://localhost:8000/answer?q=What+models+does+Groq+support"

Best settings: Default (5 pages, temp 0.3)

Good for: Implementation guides, step-by-step instructions

curl "http://localhost:8000/answer?q=How+do+I+implement+streaming"

Best settings: More pages (6-8), low temp (0.2-0.3)

Good for: Requesting code snippets

curl "http://localhost:8000/answer?q=Show+me+a+Python+example+of+calling+the+API"

Best settings: Code-focused strategy (when available), temp 0.3

Good for: Comparing models, features, approaches

curl "http://localhost:8000/answer?q=Difference+between+Llama+3.1+and+3.3"

Best settings: More pages (7-10), temp 0.3

Good for: Multi-part questions requiring reasoning

curl "http://localhost:8000/answer?q=How+to+build+a+chatbot+with+streaming+and+conversation+history"

Best settings: More pages (8-10), 70B+ model, temp 0.4

Problem: Returns error or empty answer

Possible causes:

• No search results found
• GROQ_API_KEY not set
• Rate limit exceeded

Solution:

# Check if search works
curl "http://localhost:8000/search?q=your+question"

# Verify API key
echo $GROQ_API_KEY

# Check strategy info
curl "http://localhost:8000/answer/info"

Problem: Answer is vague, incorrect, or unhelpful

Possible causes:

• Search results not relevant
• Not enough context (too few pages)
• Temperature too high

Solutions:

# Check search results first
curl "http://localhost:8000/search?q=your+question&limit=10"

# Increase context
curl "http://localhost:8000/answer?q=your+question&maxContextPages=8"

# Lower temperature
curl "http://localhost:8000/answer?q=your+question&temperature=0.2"

# Increase search quality threshold
curl "http://localhost:8000/answer?q=your+question&minScore=60"

Problem: Takes >5 seconds to respond

Possible causes:

• Slow search strategy
• Too many pages in context
• Large model

Solutions:

# Use fewer pages
curl "http://localhost:8000/answer?q=your+question&maxContextPages=3"

# Switch to faster model
curl "http://localhost:8000/answer?q=your+question&model=llama-3.1-8b-instant"

# Check timing breakdown in response metadata

cp answer/llama-3.3-70b-default.ts answer/my-custom-strategy.ts

// Change these constants
const DEFAULT_MODEL = "your-preferred-model";
const DEFAULT_MAX_CONTEXT_PAGES = 3; // Adjust as needed
const DEFAULT_TEMPERATURE = 0.5;
const MAX_TOKENS_PER_PAGE = 1000; // Adjust as needed

import { DEFAULT_SYSTEM_PROMPT } from "./utils.ts";

// Or create your own:
const CUSTOM_SYSTEM_PROMPT = `You are a helpful assistant that...`;

export const answerStrategy: AnswerStrategy = {
  name: "my-custom-strategy",
  description: "What makes your strategy special",
  answer: async (query: string, options: AnswerOptions = {}) => {
    // Your implementation
  },
};

Edit answer/index.ts:

import { answerStrategy } from "./my-custom-strategy.ts";

curl "http://localhost:8000/answer/info"  # Verify active strategy
curl "http://localhost:8000/answer?q=test+question"

1. Use faster models (8B instead of 70B)
2. Reduce context pages (2-3 instead of 5)
3. Use faster search strategy
4. Lower token limits per page

1. Use larger models (70B+)
2. Increase context pages (7-10)
3. Use higher search score thresholds
4. Include more search results to choose from

1. Start with default (5 pages, 70B, temp 0.3)
2. Monitor timing breakdown in responses
3. Adjust based on question complexity
4. Cache common questions (future enhancement)

1. ✅ Test the default strategy
2. ✅ Try different query parameters
3. ✅ Check the README.md for more details
4. ⚡ Create a fast strategy for simple questions
5. 📚 Create an extended strategy for complex questions
6. 🔧 Customize system prompts for your use case
7. 📊 Build a testing harness (like /search/test)

• Search Strategies: See /search/ for how document retrieval works
• Groq Models: https://console.groq.com/docs/models
• API Reference: https://console.groq.com/docs/api-reference

Check out:

• answer/README.md - Detailed documentation
• answer/types.ts - Type definitions
• answer/utils.ts - Helper functions
• main.tsx - How endpoints are implemented