Stage 5: Advanced Retrieval (RAG)

Overview

Stage 5 uses Retrieval Augmented Generation (RAG) to find and rank the most relevant information from all available sources, ensuring AI responses are grounded in actual data.

Processing Time: 400-1000ms
Type: Critical path - Information retrieval
Purpose: Find the best information to include in the response

---

What Happens in This Stage

RAG Pipeline

---

Three-Phase Process

Phase 1: Initial Retrieval

Data Sources:

• Knowledge Articles: Curated documentation
• Case Resolutions: Historical solutions
• Policy Documents: Official guidelines
• Salesforce Records: CRM data
• Vector Embeddings: Semantic search results
• External APIs: Third-party data

Retrieval Methods:

retrievalSources = [
  {
    source: "Knowledge Base",
    method: "keyword_search",
    results: 30
  },
  {
    source: "Vector Store",
    method: "semantic_search",
    results: 20
  },
  {
    source: "Salesforce",
    method: "soql_query",
    results: 15
  }
]

Phase 2: Re-Ranking

Scoring Criteria:

Factor	Weight	Description
Semantic Similarity	40%	How well content matches query meaning
Recency	20%	How recent the information is
Authority	15%	Source credibility and trust score
User Preferences	10%	Historical user interactions
Completeness	10%	How fully it answers the question
Accessibility	5%	User permissions and access rights

Example Re-Ranking:

initialResults = [
  {content: "Return Policy v3.2", score: 0.75},
  {content: "How to Process Returns", score: 0.82},
  {content: "Return FAQ", score: 0.68}
];

rerankedResults = applyReranking(initialResults, {
  query: "How do I process a return?",
  userPreferences: {...},
  recencyBoost: 0.1,
  authorityBoost: 0.05
});

// Results after re-ranking:
[
  {content: "Return Policy v3.2", score: 0.95},      // Boosted by authority
  {content: "How to Process Returns", score: 0.92},  // Highest semantic match
  {content: "Return FAQ", score: 0.88}               // Boosted by recency
]

Phase 3: Context Selection

Selection Strategy:

• Choose top N results (typically 3-5)
• Ensure diversity of perspectives
• Stay within token limits
• Prioritize actionable information

tokenBudget = 4000; // Available tokens for context
selectedContext = [];
usedTokens = 0;

for (result of rankedResults) {
  if (usedTokens + result.tokens < tokenBudget) {
    selectedContext.push(result);
    usedTokens += result.tokens;
  }
}

---

Retrieval Techniques

info

The following examples are conceptual illustrations of how RAG retrieval works in Agentforce. These are not actual APIs you can call directly.

1. Keyword Search

Traditional text matching example:

SELECT Title, Body, Score
FROM Knowledge__kav
WHERE Body LIKE '%return%' 
  AND Body LIKE '%process%'
  AND PublishStatus = 'Online'
ORDER BY LastModifiedDate DESC
LIMIT 30

2. Semantic Search (Vector Embeddings)

Meaning-based retrieval concept:

// Conceptual: Convert query to vector embedding
queryVector = embedModel.encode("How do I process a return?");
// [0.21, -0.15, 0.82, ..., 0.45] (768 dimensions)

// Conceptual: Find similar vectors in database
similarDocuments = vectorStore.search(queryVector, {
  topK: 20,
  threshold: 0.7,
  metric: "cosine_similarity"
});

3. Hybrid Search

Conceptual combination of keyword + semantic:

keywordResults = keywordSearch(query);      // 30 results
semanticResults = semanticSearch(query);     // 20 results

// Merge and deduplicate
combinedResults = merge(keywordResults, semanticResults);

// Re-rank the combined set
finalResults = rerank(combinedResults);

---

How Agentforce Uses RAG

Citation Tracking

Agentforce automatically tracks where information comes from:

// Conceptual response structure
response = {
  answer: "To process a return...",
  citations: [
    {
      source: "Return Policy v3.2",
      url: "https://kb.company.com/returns/policy",
      relevanceScore: 0.95,
      excerpt: "Returns must be initiated within 30 days..."
    },
    {
      source: "Case #12345",
      url: "https://company.my.salesforce.com/500xxx",
      relevanceScore: 0.88,
      excerpt: "Customer successfully returned..."
    }
  ]
}

Multi-Hop Reasoning

Agentforce can chain multiple retrievals together:

// Conceptual multi-hop retrieval
query = "What was the resolution for the last return issue from Acme Corp?"

hop1 = retrieve("Recent cases for Acme Corp");
// Found: Case #12345 - Return issue

hop2 = retrieve("Resolution details for Case #12345");
// Found: Refund processed, customer satisfied

hop3 = retrieve("Similar resolutions");
// Found: 5 similar cases with same pattern

Temporal Awareness

Recent information is prioritized:

// Conceptual recency scoring adjustments
scoringAdjustment = {
  lastWeek: +0.20,
  lastMonth: +0.10,
  lastQuarter: +0.05,
  lastYear: +0.02,
  older: -0.10
}

---

Performance Benefits

Why RAG + Re-ranking Matters

Hybrid Search Advantages:

• ✅ Combines keyword precision with semantic understanding
• ✅ Finds relevant content even when exact terms don't match
• ✅ Adapts to different query phrasings
• ✅ Respects data permissions and recency

Re-ranking Impact:

• ✅ Prioritizes most relevant results
• ✅ Balances multiple relevance factors
• ✅ Reduces irrelevant responses
• ✅ Improves user satisfaction

tip

Hybrid search (keyword + semantic) consistently outperforms either method alone, especially for complex or ambiguous queries.

---

Monitoring

Performance Considerations

Typical Retrieval Performance:

• Processing Time: 400-1000ms per request
• Sources Queried: 3-6 data sources (Knowledge, Salesforce Records, Vector Search)
• Candidates Retrieved: 30-50 initial candidates
• Final Selection: 3-5 most relevant items
• Relevance Threshold: Minimum 0.7 semantic similarity score

note

Monitor your agent's performance through Setup → Einstein → Einstein for Service. Salesforce provides built-in analytics for response quality and retrieval effectiveness.

---

Configuration

Typical Retrieval Settings

info

The following configurations represent typical values used by Agentforce's RAG engine. These settings are managed by Salesforce and are NOT user-configurable in standard Agentforce implementations. They are shown here for understanding how the system operates internally.

Example Retrieval Configuration:

{
  "retrieval": {
    "maxCandidates": 50,
    "topK": 5,
    "semanticThreshold": 0.7,
    "enableHybridSearch": true,
    "sources": {
      "knowledgeBase": {
        "enabled": true,
        "weight": 1.0
      },
      "vectorSearch": {
        "enabled": true,
        "weight": 1.2
      },
      "salesforceRecords": {
        "enabled": true,
        "weight": 0.8
      }
    }
  }
}

Typical Re-ranking Weights

Example Re-ranking Configuration:

{
  "reranking": {
    "semanticSimilarity": 0.40,
    "recency": 0.20,
    "authority": 0.15,
    "userPreferences": 0.10,
    "completeness": 0.10,
    "accessibility": 0.05
  }
}

What You CAN Configure

While the RAG engine parameters above are automatic, you can configure:

• Data Sources: Enable/disable Knowledge Base, Data Cloud, Salesforce objects
• Knowledge Articles: Select which articles are accessible
• Record Access: Configure object and field-level security
• Topic Filters: Define which topics trigger retrieval

Configure these in Setup → Agent → Topics & Actions

---

Best Practices

✅ Do's

• ✅ Use hybrid search (keyword + semantic)
• ✅ Re-rank results for optimal relevance
• ✅ Track citations for transparency
• ✅ Respect token limits in context selection
• ✅ Monitor relevance scores

❌ Don'ts

• ❌ Rely solely on keyword search
• ❌ Skip re-ranking (uses first results)
• ❌ Exceed token budgets
• ❌ Ignore user permissions
• ❌ Return uncited information

---

Troubleshooting

Issue: Low Relevance Scores

Symptoms:

• Users report irrelevant information
• Low satisfaction scores

Solutions:

1. Tune re-ranking weights
2. Expand vector embedding model
3. Add more data sources
4. Improve query understanding

Issue: Slow Retrieval

Symptoms:

• Stage 5 takes > 1000ms
• Timeout errors

Solutions:

1. Add caching layer
2. Reduce max candidates
3. Optimize vector search indices
4. Use parallel retrieval

---

Integration with Stage 6

Retrieved information moves to Stage 6: Quality & Safety Gates:

---

Related Documentation

• Stage 4: Query Planning - Previous stage
• Stage 6: Quality & Safety - Next stage
• Complete Pipeline - All stages overview

---

Stage 5 ensures AI responses are grounded in relevant, authoritative information through advanced RAG techniques, delivering 2x better relevance than basic keyword search.