Redwood - Standard RAG

Redwood is the fastest RAG strategy, using direct vector search without any prompt rewriting. It's optimized for speed and cost-effectiveness.

Overview

Redwood uses the simplest, most straightforward RAG approach:

• Takes the user's original query as-is

• Converts directly to vector embedding

• Retrieves top matching documents

• Generates response with retrieved context

Performance: ~1-2 seconds Ideal for: Clear, well-formed questions

How Redwood Works

Processing Flow

User Query: "What is the product pricing?"
     ↓
[1] Embed Query → Vector [0.12, -0.45, 0.78, ...]
     ↓
[2] Vector Search (Pinecone/TigrisDB)
     ↓
[3] Retrieve Top 5-10 Documents
     ↓
[4] Build Context from Documents
     ↓
[5] LLM Completion (with context)
     ↓
Response: "Our pricing plans are..."

Technical Details

Step 1: Query Embedding

• Model: text-embedding-ada-002

• No preprocessing or rewriting

• Original user text is embedded directly

Step 2: Vector Search

• Searches namespace: org-{orgId}

• Returns topK results (default: 5-10)

• Cosine similarity ranking

• Filters by metadata (if configured)

Step 3: Context Building

• Retrieved documents formatted as context

• Includes title, content, and URL

• Preserves source citations

• Respects max context window

Step 4: LLM Completion

• Model: GPT-4o, GPT-4, or GPT-3.5-turbo

• System prompt + context + user query

• Temperature: 0.7 (default)

• Streams response (optional)

Performance Characteristics

Latency Breakdown

Query Embedding:      ~100ms
Vector Search:        ~200ms
Context Building:     ~50ms
LLM Completion:       ~800ms
Response Streaming:   ~200ms
────────────────────────────
Total:                ~1.4s

Token Usage

Component	Tokens	Notes
System Prompt	150-300	Agent instructions
Retrieved Context	800-1500	Top 5-10 documents
User Query	10-50	Original question
Response	150-400	Generated answer
Total	~1,500-2,000	Per request

Cost Implications

Per 1,000 Requests (GPT-3.5-turbo):

• Embedding: ~$0.01

• LLM Completion: ~$0.30

• Vector Search: ~$0.05

• Total: ~$0.36

When to Use Redwood

✅ Ideal Use Cases

1. FAQ Bots

Q: "What are your business hours?"
→ Direct, clear question
→ No context needed
→ Fast response critical

2. Product Information Lookup

Q: "Does the Pro plan include SSO?"
→ Straightforward fact query
→ Answer is in docs
→ Speed matters

3. Quick Reference Tools

Q: "How do I export a CSV report?"
→ Clear task-based question
→ Step-by-step answer in docs
→ Users want fast help

4. API Documentation Queries

Q: "What parameters does the /users endpoint accept?"
→ Technical, precise question
→ Documented information
→ Developers value speed

❌ Not Ideal For

1. Ambiguous Questions

Q: "How much does it cost?"
→ Missing context: which product? which plan?
→ Redwood can't infer missing information
→ Cedar or Cypress would rewrite to clarify

2. Follow-up Questions

User: "What are your pricing plans?"
Agent: "We have Free, Pro, and Enterprise..."
User: "What's included in that one?" ❌
→ "that one" requires conversation context
→ Cedar would rewrite: "What's included in the Enterprise plan?"

3. Complex Multi-Part Queries

Q: "I need to migrate from Competitor X, what's the process and will my data transfer automatically?"
→ Multiple sub-questions
→ Requires breaking down and context
→ Cypress would handle better

Configuration

Agent Settings

When using Redwood strategy:

{
  "strategyCode": "STANDARD_RAG",
  "topK": 5-10,              // Number of documents to retrieve
  "temperature": 0.7,         // LLM creativity
  "maxTokens": 500,           // Response length limit
  "model": "gpt-3.5-turbo"    // Or gpt-4o for higher quality
}

Optimization Tips

1. Optimize topK

Too low (3):   May miss relevant context
Sweet spot (5-10): Balance of context and speed
Too high (20+): Slower, noisier context

2. Document Quality

• Well-structured source documents = better retrieval

• Clear headings and sections

• Avoid very long documents (chunk effectively)

3. Query Quality Training

• Educate users to ask clear questions

• Provide example questions

• Use suggested prompts

Comparison with Other Strategies

vs. Cedar (Context-Aware)

Redwood Advantages:

• ⚡ Faster (~1s faster than Cedar)

• 💰 Cheaper (one less LLM call)

• 📊 Simpler to debug

Cedar Advantages:

• 🧠 Better for conversational queries

• 🔄 Handles follow-ups better

• 📝 Clarifies ambiguous questions

Example:

Query: "What about pricing?"

Redwood: Searches for "what about pricing"
→ May return generic results

Cedar: Rewrites to "What are the pricing plans for [Product X]?"
→ More targeted results

vs. Cypress (Advanced)

Redwood Advantages:

• ⚡⚡ Much faster (~2-3s faster)

• 💰💰 Much cheaper (no reranking, expansion)

• 🎯 Simpler implementation

Cypress Advantages:

• 🎯 Higher accuracy

• 🔍 Better semantic matching through query expansion

• 📊 Tier-based source organization

• 🏆 Reranking improves precision

Example:

Query: "password reset"

Redwood: Searches for "password reset"
→ Finds exact matches

Cypress: Expands to "password reset, change password, 
         recover account, reset credentials, account recovery"
→ Finds more variations
→ Reranks to top 10 best matches

Real-World Performance

Case Study: Developer Documentation Site

Setup:

• 5,000 API documentation pages

• Average query: "How to use [endpoint]"

• 10,000 queries/day

Redwood Performance:

• Average latency: 1.2 seconds

• 95th percentile: 1.8 seconds

• User satisfaction: 4.2/5

• Cost: $3.60/day

Result: Perfect fit for clear, technical queries

Case Study: E-commerce FAQ

Setup:

• 500 FAQ articles

• Average query: "What is [policy]?"

• 5,000 queries/day

Redwood Performance:

• Average latency: 1.0 seconds

• 95th percentile: 1.5 seconds

• User satisfaction: 4.5/5

• Cost: $1.80/day

Result: Fast, accurate for straightforward questions

Monitoring Redwood

Key Metrics to Track

1. Response Time

Target: < 2 seconds (95th percentile)
Alert: > 3 seconds

2. Retrieval Quality

Metric: Citation relevance rate
Target: > 85% of responses have relevant citations

3. Answer Rate

Metric: % of queries with an answer
Target: > 90%

4. Cost per Query

Track: Token usage trends
Optimize: Adjust topK, max response tokens

Common Issues

Slow Responses:

• Check vector DB latency

• Verify network connectivity

• Consider caching frequent queries

Irrelevant Results:

• Improve document chunking

• Add metadata filters

• Consider switching to Cedar for ambiguous queries

Low Answer Rate:

• Ensure knowledge base has sufficient coverage

• Check data source connectivity

• Review unanswered queries for patterns

Best Practices

1. Document Preparation

✅ Clear, well-structured documents ✅ Good titles and headings ✅ Logical chunking (200-500 tokens) ✅ Updated regularly

2. User Guidance

✅ Provide example questions ✅ Show suggested prompts ✅ Educate on asking clear questions

3. Performance Optimization

✅ Monitor latency metrics ✅ Cache common queries ✅ Use appropriate topK ✅ Right-size context window

4. Quality Assurance

✅ Regular testing with sample queries ✅ Review low-confidence responses ✅ A/B test against Cedar for borderline cases

Migration Path

When to Switch from Redwood

Switch to Cedar if:

• 30% of queries are follow-ups

• Users report ambiguous results

• Conversational use increases

Switch to Cypress if:

• Accuracy is more important than speed

• Budget allows for higher costs

• Query complexity increases significantly

Code Example

Using Redwood via API

curl -X POST https://api.twig.so/api/chat \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "prompt": "What is the refund policy?",
    "agentId": "agent-123",
    "strategyCode": "STANDARD_RAG",
    "topK": 5
  }'

Response Format

{
  "response": "Our refund policy allows returns within 30 days...",
  "sources": [
    {
      "title": "Refund Policy",
      "url": "https://example.com/refund-policy",
      "relevance": 0.95
    }
  ],
  "latency": 1.2,
  "tokensUsed": 1523,
  "strategy": "STANDARD_RAG"
}

Next Steps

• Cedar Strategy - When to upgrade for conversational queries

• Cypress Strategy - Maximum accuracy with reranking

• Performance Tuning

• Cost Optimization