Redwood vs Cedar Decision

The Problem

Unclear when to use Redwood (standard RAG) versus Cedar (context-aware RAG), causing agents to choose incorrectly and miss user-specific context.

Symptoms

• ❌ Generic answer when user-specific needed

• ❌ Cedar used unnecessarily (added latency)

• ❌ Cannot explain Redwood vs Cedar choice

• ❌ Inconsistent strategy for similar queries

• ❌ Misses user permissions/context

Real-World Example

Query: "What documents can I access?"

Redwood (wrong choice):
→ Returns: "All users can access help docs, API docs..."
→ Generic answer, ignores "I"

Cedar (correct choice):
→ Retrieves user context: User = Alice, Team = Engineering
→ Returns: "You (Alice) can access: Engineering docs, Prod database docs,
shared team folders..."
→ Personalized, permission-aware

Agent chose Redwood: Missed user-specific context

---

Deep Technical Analysis

Redwood Characteristics

Use Cases:

✓ Factual queries
✓ Public documentation
✓ No user context needed
✓ Fast, simple retrieval

Examples:
→ "What is the API rate limit?"
→ "How to install the SDK?"
→ "Explain authentication flow"

Speed: 1-2 seconds
Cost: Low

Cedar Characteristics

Use Cases:

✓ User-specific queries
✓ Permission-based access
✓ Contextual answers
✓ Personalized responses

Examples:
→ "What can I access?"
→ "My recent projects"
→ "Documents for my team"

Speed: 2-4 seconds (retrieves user context first)
Cost: Medium (additional context retrieval)

Decision Criteria

Query Analysis:

Personal pronouns:
→ "I", "my", "me", "our" → Cedar

Possession queries:
→ "What do I have?"
→ "My permissions" → Cedar

Role-based:
→ "As a developer, what..."
→ "For my role..." → Cedar

Generic:
→ "What is X?"
→ "How does Y work?" → Redwood

User Context Indicators:

Check if query needs:
1. User identity → Cedar
2. User permissions → Cedar
3. User's team/group → Cedar
4. User's previous actions → Cedar

If none: Use Redwood

Hybrid Approach

Context Pre-Check:

Algorithm:
1. Classify query: Personal vs Generic
2. If personal:
   a. Fetch user context (team, role, permissions)
   b. Use Cedar with context
3. If generic:
   a. Use Redwood (no context needed)

Efficient: Only fetch context when needed

Progressive Enhancement:

Start Redwood:
→ Retrieve generic docs
→ Check if answer sufficient

If insufficient:
→ "This answer applies generally. Would you like personalized info?"
→ User confirms
→ Switch to Cedar, add context

User-controlled escalation

Context Sources (Cedar)

User Metadata:

Retrieve from auth/user DB:
→ user_id: "12345"
→ name: "Alice Johnson"
→ email: "[email protected]"
→ role: "Senior Engineer"
→ team: "Backend Team"
→ permissions: ["read_prod_db", "deploy_staging"]

Activity History:

Recent user actions:
→ Last accessed: Project Phoenix docs
→ Frequently queries: API authentication
→ Team: Working on OAuth integration

Context: Bias retrieval toward OAuth/auth topics

---

How to Solve

Detect personal pronouns (I, my, me, our) → use Cedar + identify permission/access queries → use Cedar + use Redwood by default for generic queries + fetch user context only when Cedar selected + measure Cedar accuracy (correct personalization?) + monitor Cedar usage rate (over/under-use) + allow user to toggle "personalized mode" explicitly + test with user-specific eval queries. See Redwood vs Cedar.