Vector Search & Embeddings

Overview

Embeddings are the core technology that powers semantic search in RAG systems. They transform text into high-dimensional vectors that capture meaning, enabling your system to find relevant information based on conceptual similarity rather than just keyword matching. However, embeddings and vector search introduce their own set of challenges that can severely impact retrieval quality.

Why Vector Search Matters

Effective vector search enables:

• Semantic understanding - Find conceptually similar content, not just exact matches
• Multilingual retrieval - Match queries and documents across languages
• Robust search - Handle typos, synonyms, and paraphrasing naturally
• Contextual relevance - Retrieve based on meaning and intent

Poor vector search results in:

• Retrieval failures - Relevant content exists but isn't found
• Irrelevant results - Documents returned have high similarity scores but wrong context
• Inconsistent quality - Search works for some queries but fails for others
• Degraded performance over time - Embedding drift as models or data changes

Common Vector Search Challenges

Embedding Quality

• Poor semantic search results - Wrong documents ranked highly
• Embedding model drift - Performance degrades after model updates
• Domain-specific vocabulary - General embeddings miss specialized terms
• Multilingual issues - Cross-language retrieval fails

Index Management

• Vector index out of sync - Embeddings don't match current documents
• Dimensionality mismatch - Incompatible embedding dimensions
• Cold start problem - Insufficient data for quality embeddings
• Performance degradation - Slow queries as index grows

Scoring & Calibration

• Similarity score calibration - Threshold tuning and interpretation
• Inconsistent similarity scores - Scores not comparable across queries
• False positives/negatives - Wrong confidence in retrieval results

Cost & Performance

• Embedding cost optimization - Balancing quality with API costs
• Vector database performance - Query latency and throughput issues
• Scale challenges - Performance at millions of vectors

Solutions in This Section

Browse these guides to optimize your vector search:

• Poor Semantic Search Results
• Embedding Model Drift
• Cold Start Problem
• Vector Index Out of Sync
• Dimensionality Mismatch
• Similarity Score Calibration
• Multilingual Embedding Issues
• Domain-Specific Vocabulary
• Embedding Cost Optimization
• Vector Database Performance

Embedding Models: Choosing the Right One

Different embedding models have different strengths:

Model Type	Use Case	Pros	Cons
General-purpose (OpenAI, Cohere)	Broad knowledge domains	Great out-of-box performance	May miss domain-specific terms
Multilingual (mBERT, LaBSE)	Cross-language retrieval	Language-agnostic search	Lower performance per-language
Domain-specific	Legal, medical, technical	High accuracy in domain	Poor generalization outside domain
Lightweight	Cost-sensitive, high-volume	Low latency, low cost	Reduced semantic understanding

Key decision factors:

• Domain specialization needs
• Language requirements
• Query volume and cost constraints
• Latency requirements
• Customization needs (fine-tuning capability)

Best Practices

Embedding Strategy

1. Match model to use case - Domain-specific vs general-purpose
2. Consistent embedding - Use same model for queries and documents
3. Version control - Track which embedding model created which vectors
4. Test before switching - Evaluate impact of model changes on retrieval quality

Index Management

1. Keep index synchronized - Re-embed when documents change
2. Monitor index health - Track index size, query latency, recall rates
3. Implement fallback strategies - Hybrid search (vector + keyword)
4. Optimize for scale - Use appropriate vector DB and index types

Quality Assurance

1. Calibrate similarity thresholds - Determine meaningful score ranges
2. Validate retrieval quality - Regular testing with representative queries
3. Monitor drift - Track retrieval metrics over time
4. A/B test changes - Compare embedding models and strategies

Cost Optimization

1. Batch embedding operations - Reduce API calls
2. Cache embeddings - Don't re-embed unchanged content
3. Use tiered models - Expensive models for queries, cheaper for documents
4. Consider open-source - Self-hosted models for high-volume use cases

Impact on RAG Performance

Vector search quality has cascading effects:

Poor Embeddings
    ↓
Wrong Documents Retrieved
    ↓
Irrelevant Context to LLM
    ↓
Hallucinated or Incorrect Answers
    ↓
Loss of User Trust

Stage	Impact of Good Embeddings	Impact of Bad Embeddings
Retrieval	Relevant documents found	Wrong or missing documents
Ranking	Best documents ranked first	Irrelevant docs ranked highly
Context	High-quality input to LLM	Poor or misleading context
Answers	Accurate, grounded responses	Hallucinations and errors
User Experience	Trust and satisfaction	Frustration and abandonment

Advanced Techniques

Hybrid Search

Combine vector search with other retrieval methods:

• Vector + Keyword - Semantic understanding + exact matching
• Vector + Filters - Semantic search within metadata constraints
• Multi-stage retrieval - Broad vector search → precise reranking

Query Enhancement

Improve retrieval by transforming queries:

• Query expansion - Add synonyms and related terms
• Hypothetical document embeddings (HyDE) - Embed generated answer, not question
• Multi-query strategies - Generate and search multiple query variations

Context Enrichment

Enhance document embeddings with:

• Metadata augmentation - Include title, tags, source in embedding
• Hierarchical context - Embed documents with parent section context
• Cross-references - Link related documents in vector space

Reranking

Post-process retrieval results:

1. Vector search retrieves candidate documents (top 20-50)
2. Reranker (cross-encoder) scores candidates with query
3. Return top reranked results (top 5-10)

This two-stage approach balances speed and accuracy.

Quick Diagnostics

Signs your embeddings need attention:

• ✗ Searches return obviously irrelevant documents
• ✗ Known relevant documents aren't retrieved
• ✗ Similarity scores are clustered (all high or all low)
• ✗ Specialized terms don't retrieve correct documents
• ✗ Retrieval quality varies wildly across query types
• ✗ Performance degraded after a model update

Signs your embeddings are working well:

• ✓ Semantically similar queries return similar documents
• ✓ Similarity scores correlate with human relevance judgments
• ✓ Retrieval handles synonyms and paraphrasing
• ✓ Cross-lingual queries work (if needed)
• ✓ Domain-specific terms retrieve correctly
• ✓ Consistent performance across query types

Monitoring & Metrics

Track these metrics to ensure embedding quality:

Retrieval Metrics

• Recall@k - What % of relevant docs are in top k results?
• Precision@k - What % of top k results are relevant?
• MRR (Mean Reciprocal Rank) - How quickly do relevant docs appear?
• NDCG (Normalized Discounted Cumulative Gain) - Quality of ranking

Operational Metrics

• Query latency - P50, P95, P99 response times
• Index size - Vector count and storage requirements
• Embedding cost - API costs for embedding generation
• Refresh lag - Time between content update and re-embedding

Quality Metrics

• Similarity score distribution - Are scores well-calibrated?
• Retrieval diversity - Are results too similar to each other?
• Coverage - How much of knowledge base is being retrieved?

Bottom line: Embeddings are the "eyes" of your RAG system. If they don't capture meaning accurately, everything downstream suffers. Invest time in getting this right.

---

Agent Instructions: Querying This Documentation

If you need additional information that is not directly available in this page, you can query the documentation dynamically by asking a question.

Perform an HTTP GET request on the current page URL with the ask query parameter:

GET /dev/rag-scenarios-and-solutions/vectors.md?ask=<question>

The question should be specific, self-contained, and written in natural language. The response will contain a direct answer to the question and relevant excerpts and sources from the documentation.

Use this mechanism when the answer is not explicitly present in the current page, you need clarification or additional context, or you want to retrieve related documentation sections.