14 — Similarity Search Techniques¶

Learn how vector similarity search retrieves semantically relevant information from high-dimensional embeddings and understand the algorithms, distance metrics, filtering strategies, ranking approaches, and production trade-offs behind modern semantic retrieval systems.

📖 Overview¶

Similarity search is the core retrieval operation that allows a RAG system to find content that is semantically related to a user query.

The basic workflow is:

User Query
    ↓
Query Embedding
    ↓
Vector
    ↓
Similarity Search
    ↓
Candidate Vectors
    ↓
Top-K Results
    ↓
Retrieved Context

The goal is simple:

Given a query vector, find the most relevant vectors from a large collection.

For example:

Query:

"What is the employee leave entitlement?"

             ↓
        Query Embedding

             ↓

      Vector Similarity Search

             ↓

┌─────────────────────────────────┐
│ Annual Leave Policy              │
│ Employee Leave Entitlement      │
│ Leave Carry-Forward Rules       │
└─────────────────────────────────┘

Similarity search is therefore the bridge between:

Embeddings

and:

Retrieval

1. Why Similarity Search Matters¶

An embedding converts content into a numerical representation.

For example:

"Annual leave entitlement"

might become:

[0.12, -0.45, 0.77, 0.21, ...]

A different piece of text:

"Employees receive 25 days of paid leave"

may produce a vector located relatively close to the query vector.

Similarity search identifies these nearby vectors.

                 Vector Space

                    ● Query
                   / \
                  /   \
                 ●     ●
              Relevant  Relevant

       ●
   Unrelated

The closer the vectors are according to the selected similarity metric, the more likely the content is to be semantically related.

2. Similarity Search in RAG¶

Similarity search is part of the retrieval stage.

flowchart LR
    A["Documents"] --> B["Chunking"]
    B --> C["Embedding"]
    C --> D["Vector Store"]

    E["User Query"] --> F["Query Embedding"]
    F --> D

    D --> G["Similarity Search"]
    G --> H["Top-K Chunks"]
    H --> I["Context Assembly"]
    I --> J["LLM"]

The quality of similarity search directly affects the quality of retrieved context.

3. Query Embedding¶

The first step is converting the user query into the same vector space used by the stored documents.

Stored Documents
      ↓
Embedding Model
      ↓
Document Vectors

User Query
      ↓
Same Embedding Model
      ↓
Query Vector

The query and document embeddings must be compatible.

4. Same Embedding Space¶

Suppose documents were embedded using:

Embedding Model A
Dimension = 768

The query should normally be embedded using the same compatible model:

Embedding Model A
Dimension = 768

Avoid:

Documents → Model A
Query     → Model B

unless the models are explicitly designed to work together.

5. Vector Space¶

A vector database can be thought of as storing points in a high-dimensional space.

For visualization:

        y
        ↑
        │        ● Document A
        │
        │   ● Query
        │
        │             ● Document B
        │
        └────────────────────→ x

Real embedding spaces may contain:

or more dimensions.

Humans cannot directly visualize these dimensions, but similarity algorithms can operate on them efficiently.

6. Similarity vs Distance¶

Two broad approaches are common.

Similarity¶

Higher value means:

More Similar

Examples:

Cosine Similarity
Dot Product

Distance¶

Lower value means:

More Similar

Examples:

Euclidean Distance
Manhattan Distance

The vector database must know which interpretation to use.

7. Cosine Similarity¶

Cosine similarity measures the angle between two vectors.

Conceptually:

              Vector A
                ↗
               /
              / θ
             /
            ●────────────→ Vector B

The key idea is:

Smaller angle
     ↓
Higher similarity

Cosine similarity is widely used for text embeddings.

8. Cosine Similarity Formula¶

Conceptually:

cosine similarity =
(dot product of A and B)
/
(product of their magnitudes)

In mathematical notation:

cos(θ) = (A · B) / (||A|| ||B||)

For normalized vectors, the relationship between cosine similarity and dot product becomes especially convenient.

9. Dot Product¶

The dot product measures vector alignment.

For:

A = [a1, a2, a3]

B = [b1, b2, b3]

the dot product is:

A · B =
a1b1 + a2b2 + a3b3

For normalized embeddings, dot-product search is commonly used as an efficient similarity measure.

10. Euclidean Distance¶

Euclidean distance measures straight-line distance between vectors.

Conceptually:

A ●
   \
    \
     \
      ● B

Smaller distance means:

A and B are closer

For high-dimensional embedding spaces, whether Euclidean distance is appropriate depends on the embedding model and vector normalization.

11. Similarity Metric Comparison¶

Metric	Better Result	Common Use
Cosine Similarity	Higher	Text embeddings
Dot Product	Higher	Normalized embeddings / retrieval
Euclidean Distance	Lower	Geometric distance
Manhattan Distance	Lower	Specific vector workloads

The metric should be selected based on the embedding model and retrieval behavior rather than personal preference.

12. Choosing the Similarity Metric¶

A practical decision process is:

flowchart TD
    A["Embedding Model"] --> B["Check Model Guidance"]
    B --> C["Check Normalization"]
    C --> D["Select Metric"]
    D --> E["Build Index"]
    E --> F["Evaluate Retrieval"]
    F --> G["Validate Recall + Latency"]

Do not choose a metric independently from the embedding model.

13. Exact Nearest Neighbor Search¶

The simplest search strategy compares the query vector against every stored vector.

Query Vector
     ↓
Vector 1 → Similarity
Vector 2 → Similarity
Vector 3 → Similarity
...
Vector N → Similarity
     ↓
Sort Scores
     ↓
Top-K

This is often called:

Brute-force or exact nearest-neighbor search.

14. Exact Search Complexity¶

For:

N vectors

the system may need to compare the query against:

N vectors

for every search.

Conceptually:

O(N)

candidate comparisons.

For a small dataset this may be perfectly acceptable.

For millions or billions of vectors, it becomes increasingly expensive.

15. Exact Search Example¶

A simple Python implementation:

import numpy as np


def cosine_similarity(a, b):
    return np.dot(a, b) / (
        np.linalg.norm(a) *
        np.linalg.norm(b)
    )


def search(query, vectors, top_k=5):

    scores = [
        cosine_similarity(query, vector)
        for vector in vectors
    ]

    ranked = np.argsort(scores)[::-1]

    return ranked[:top_k]

This is useful for understanding the basic concept.

Production systems generally use optimized vector indexes rather than scanning every vector in application code.

16. Approximate Nearest Neighbor Search¶

Large-scale systems commonly use:

Approximate Nearest Neighbor (ANN) search.

Instead of searching every vector:

Millions of Vectors
        ↓
ANN Index
        ↓
Smaller Candidate Set
        ↓
Top-K Results

ANN trades a small amount of potential recall for significantly improved search performance.

17. ANN Trade-off¶

The fundamental trade-off is:

Search Accuracy
      ↕
Search Speed

Increasing search effort can improve recall:

Higher Search Effort
        ↓
Potentially Better Recall
        ↓
Higher Latency

Reducing search effort:

Lower Search Effort
        ↓
Lower Latency
        ↓
Potentially Lower Recall

18. ANN Architecture¶

flowchart TD
    A["Large Vector Collection"] --> B["ANN Index"]

    C["Query Vector"] --> B

    B --> D["Candidate Region"]
    D --> E["Candidate Vectors"]
    E --> F["Similarity Scoring"]
    F --> G["Top-K Results"]

19. HNSW¶

One of the most popular ANN techniques is:

Hierarchical Navigable Small World (HNSW).

HNSW organizes vectors into graph-like structures.

Conceptually:

Higher Layer

A ───────────── D
               │
               E


Lower Layer

A ── B ── C ── D
     │    │
     E ── F

The higher levels help navigate quickly toward promising regions.

The lower levels provide more detailed neighborhood search.

20. HNSW Search¶

Conceptually:

Query
  ↓
Start at upper layer
  ↓
Find closer node
  ↓
Move toward target region
  ↓
Descend to lower layer
  ↓
Search local neighborhood
  ↓
Return Top-K

This reduces the amount of the graph that needs to be explored.

21. HNSW Parameters¶

Common parameters include:

M
efConstruction
efSearch

M¶

Controls graph connectivity.

Higher values can improve recall but may increase:

Memory
Index Size
Construction Cost

efConstruction¶

Controls the amount of effort used while building the index.

Higher values can improve index quality but increase construction cost.

efSearch¶

Controls the amount of effort during query execution.

Higher values can improve recall but increase query latency.

22. HNSW Tuning¶

A simplified relationship:

efSearch ↑
    ↓
More Candidates Explored
    ↓
Recall ↑
    ↓
Latency ↑

Therefore:

Tune HNSW parameters using the actual production workload rather than relying on generic defaults.

23. IVF¶

Another ANN family is:

Inverted File (IVF).

IVF divides the vector space into clusters.

Vector Collection
       ↓
    Clustering
       ↓
┌──────┬──────┬──────┐
│ C1   │ C2   │ C3   │
└──────┴──────┴──────┘

At query time, the system searches only selected clusters.

24. IVF Search¶

flowchart TD
    A["All Vectors"] --> B["Cluster Assignment"]

    B --> C["Cluster 1"]
    B --> D["Cluster 2"]
    B --> E["Cluster 3"]
    B --> F["Cluster N"]

    G["Query Vector"] --> H["Find Relevant Clusters"]
    H --> C
    H --> D

    C --> I["Candidate Search"]
    D --> I

    I --> J["Top-K"]

The number of clusters examined affects the speed/recall trade-off.

25. IVF Search Parameters¶

A common concept is:

Number of Probes

Higher probing:

More Clusters Searched
       ↓
Potentially Higher Recall
       ↓
Higher Latency

Lower probing:

Fewer Clusters
       ↓
Lower Latency
       ↓
Potentially Lower Recall

26. Product Quantization¶

Product Quantization (PQ) compresses vector representations.

Conceptually:

Original Vector
      ↓
Split into Subvectors
      ↓
Quantize Each Subvector
      ↓
Compressed Representation

This can significantly reduce memory usage for very large vector collections.

27. PQ Architecture¶

flowchart LR
    A["High-Dimensional Vector"] --> B["Subvector 1"]
    A --> C["Subvector 2"]
    A --> D["Subvector 3"]
    A --> E["Subvector N"]

    B --> F["Quantization"]
    C --> F
    D --> F
    E --> F

    F --> G["Compressed Vector"]

The trade-off is:

Lower Memory
      ↕
Potentially Lower Search Accuracy

28. HNSW vs IVF¶

Characteristic	HNSW	IVF
Structure	Graph	Clusters
Search	Graph traversal	Cluster search
Tuning	efSearch / M	Probes / clusters
Memory	Can be significant	Often configurable
Dynamic Updates	Generally convenient	Depends on implementation
Large-Scale Search	Strong	Strong
Compression	Can combine with compression	Often combined with PQ

The actual performance depends heavily on the database implementation and workload.

29. HNSW + Quantization¶

Some systems combine:

HNSW
+
Vector Quantization

This can provide:

Fast Search
+
Reduced Memory

while attempting to maintain acceptable recall.

30. Search Pipeline¶

A production vector search operation can look like:

User Query
    ↓
Query Embedding
    ↓
Vector Index
    ↓
ANN Candidate Search
    ↓
Similarity Scoring
    ↓
Metadata Filtering
    ↓
Top-K
    ↓
Optional Reranking

The exact order may vary by vector database.

31. Top-K Retrieval¶

K determines the number of results returned.

For:

K = 5

the retriever returns:

Result 1
Result 2
Result 3
Result 4
Result 5

Larger K can increase recall but may also introduce noise.

32. Choosing K¶

A useful principle:

K too small
    ↓
Relevant information may be missed

while:

K too large
    ↓
More irrelevant context
    ↓
Higher latency / cost

Therefore K should be evaluated using the target query set.

33. Similarity Threshold¶

A retriever can also apply a minimum similarity threshold.

Conceptually:

Query
  ↓
Search
  ↓
Results
  ↓
Similarity >= Threshold
  ↓
Accepted Results

For example:

Top-K = 10
Threshold = configured minimum

Only sufficiently relevant results are retained.

34. Why Thresholds Are Difficult¶

Similarity scores depend on:

Embedding Model
Similarity Metric
Vector Normalization
Corpus
Query Type
Index

Therefore:

0.80

does not universally mean:

Highly Relevant

across all systems.

Thresholds should be calibrated experimentally.

35. Metadata Filtering¶

Similarity search can be combined with structured filters.

Example:

Query:
"What is the leave policy?"

Filter:
country = India
department = HR

Conceptually:

Semantic Similarity
        +
Metadata Constraints
        ↓
Relevant Authorized Results

36. Pre-Filtering¶

A system may first restrict the candidate set:

All Vectors
    ↓
Metadata Filter
    ↓
Eligible Vectors
    ↓
Similarity Search

Example:

tenant_id = tenant-a

This can reduce the search space.

37. Post-Filtering¶

Another approach:

All Vectors
    ↓
Similarity Search
    ↓
Candidate Results
    ↓
Metadata Filter

This can create problems if the initial Top-K results contain many records that are later filtered out.

For example:

Top-K = 5

After security filtering:
0 results

The correct approach depends on database capabilities and retrieval requirements.

38. Filtered ANN Search¶

Production vector databases often provide mechanisms for combining:

ANN Search
+
Metadata Filtering

This is particularly important for enterprise workloads involving:

Tenant Isolation
Access Control
Department Restrictions
Document Types
Geography
Time Ranges
Security Classification

39. Hybrid Search¶

Similarity search does not always replace keyword search.

Consider:

Query:
"Policy ID HR-2026-0042"

Exact keyword matching may be extremely useful.

Semantic search may instead interpret:

HR policy identifier

Hybrid retrieval combines:

Keyword Search
+
Semantic Search

40. Hybrid Search Architecture¶

flowchart TD
    A["User Query"] --> B["Keyword Search"]
    A --> C["Vector Similarity Search"]

    B --> D["Keyword Results"]
    C --> E["Semantic Results"]

    D --> F["Result Fusion"]
    E --> F

    F --> G["Combined Ranking"]
    G --> H["Top-K"]

Hybrid retrieval is particularly useful when both semantic meaning and exact terms matter.

41. Result Fusion¶

Two retrieval systems may produce:

Keyword Results

and:

Vector Results

These can be combined using ranking strategies.

Conceptually:

Keyword Score
       +
Semantic Score
       ↓
Combined Score

More advanced fusion methods are covered later in the retrieval section.

42. Reranking¶

Initial vector search may return:

Top 20 candidates

A more sophisticated reranker can then reorder them:

20 Candidates
     ↓
Reranker
     ↓
Top 5 High-Quality Results

This creates a two-stage retrieval architecture.

43. Two-Stage Retrieval¶

flowchart LR
    A["Query"] --> B["Embedding"]
    B --> C["ANN Search"]
    C --> D["Top 20 Candidates"]
    D --> E["Reranker"]
    E --> F["Top 5 Results"]

This can improve precision because:

ANN

focuses on fast candidate generation, while:

Reranker

performs deeper relevance evaluation.

Advanced reranking techniques will be covered later in the handbook.

44. Candidate Generation vs Ranking¶

It is useful to separate:

Candidate Generation¶

Goal:

Find potentially relevant content quickly.

Typical technology:

ANN Vector Search

Ranking¶

Goal:

Order candidates by relevance.

Possible technology:

Reranker
Hybrid Ranking
Business Rules

45. Similarity Search and Recall¶

Retrieval quality can be measured using:

Recall@K

The question is:

Did the relevant document or chunk appear within the top K results?

Example:

Expected relevant chunk:
Chunk 42

Retrieved:
Chunk 7
Chunk 12
Chunk 42
Chunk 81
Chunk 93

For:

K = 5

the relevant chunk was successfully retrieved.

46. Precision¶

Precision asks:

How many retrieved results are actually relevant?

Example:

Top 5 Results

Relevant:
3

Irrelevant:
2

Then:

Precision@5 = 3 / 5

Precision and recall often need to be balanced.

47. Recall vs Precision¶

Higher Recall
      ↓
Retrieve more potentially relevant content
      ↓
Potentially more noise

Higher Precision
      ↓
Retrieve more focused content
      ↓
Potentially miss relevant information

The ideal retrieval system balances both.

48. MRR¶

Mean Reciprocal Rank (MRR) focuses on the position of the first relevant result.

For one query:

Relevant result at rank 1

RR = 1 / 1
   = 1

If relevant result appears at rank 4:

RR = 1 / 4

MRR averages reciprocal rank across queries.

49. NDCG¶

Normalized Discounted Cumulative Gain (NDCG) evaluates ranking quality when multiple results may have different relevance levels.

For example:

Result 1 → Highly Relevant
Result 2 → Relevant
Result 3 → Slightly Relevant
Result 4 → Irrelevant

NDCG rewards systems that place highly relevant results near the top.

50. Similarity Search Evaluation Dataset¶

Create representative queries:

Query 1
Query 2
Query 3
...
Query N

For each query define:

Expected Relevant Chunks

Then compare:

Metric
Index
K
Threshold
Filtering

51. Search Evaluation Workflow¶

flowchart TD
    A["Evaluation Queries"] --> B["Embedding"]
    B --> C["Similarity Search"]
    C --> D["Retrieved Results"]

    E["Ground Truth"] --> F["Evaluation"]
    D --> F

    F --> G["Recall@K"]
    F --> H["Precision@K"]
    F --> I["MRR"]
    F --> J["NDCG"]

52. Benchmarking Similarity Search¶

A production benchmark should include:

Realistic Dataset Size
Representative Queries
Expected Results
Concurrent Users
Metadata Filters
Target K
Index Configuration

Measure:

Recall
Latency
Throughput
Memory
CPU
Storage

53. Latency Metrics¶

Do not measure only average latency.

Track:

P50
P95
P99

For example:

P50 = 25 ms
P95 = 70 ms
P99 = 120 ms

P95 and P99 are particularly useful for understanding tail latency.

54. Throughput¶

A production vector search system may need to support:

100 QPS
1,000 QPS
10,000 QPS

depending on workload.

Benchmark:

Queries Per Second

alongside latency.

55. Recall-Latency Trade-off¶

ANN tuning often produces:

Search Effort ↑
     ↓
Recall ↑
     ↓
Latency ↑

A production system should identify an acceptable operating point.

        Recall
          ↑
          │          ●
          │       ●
          │    ●
          │ ●
          └────────────────→
                Latency

The best point is usually not the absolute maximum recall.

56. Index Build Time¶

Vector indexes also have construction costs.

For large datasets:

Millions of vectors
       ↓
Index Construction
       ↓
CPU
Memory
Time

This matters when:

Embedding Model Changes

or:

Full Re-indexing

is required.

57. Dynamic Updates¶

Some vector workloads change frequently:

Documents Added
Documents Updated
Documents Deleted

The search architecture should therefore consider:

Index Update Cost
Consistency
Refresh Behavior
Deletion Handling

58. Real-Time vs Batch Indexing¶

Batch¶

Documents
 ↓
Batch Processing
 ↓
Batch Embedding
 ↓
Batch Indexing

Useful for:

Large Knowledge Bases
Scheduled Updates

Near Real-Time¶

Document Change
 ↓
Processing
 ↓
Embedding
 ↓
Index Update

Useful when:

Knowledge changes frequently.

59. Similarity Search in Enterprise Systems¶

Enterprise queries may require:

Semantic Similarity
+
Security
+
Tenant Isolation
+
Metadata
+
Versioning
+
Recency

Therefore a production retrieval request may look conceptually like:

{
  "query": "What is the leave policy?",
  "top_k": 10,
  "filters": {
    "tenant_id": "tenant-a",
    "department": "HR",
    "country": "IN"
  }
}

60. Security Must Precede Trust¶

A dangerous retrieval architecture is:

Search
 ↓
Retrieve Sensitive Documents
 ↓
Filter Later

Security restrictions should be enforced as early as the retrieval infrastructure allows.

Identity
 ↓
Authorization Context
 ↓
Allowed Search Space
 ↓
Similarity Search
 ↓
Results

Prompt instructions must never be treated as a security boundary.

61. Multi-Tenant Similarity Search¶

A multi-tenant system might use:

tenant_id

as a mandatory filter.

Query
 ↓
Tenant Context
 ↓
Vector Search
 ↓
Tenant-A Results

Never allow:

Tenant A Query
      ↓
Tenant B Data

to enter the LLM context.

62. Similarity Search and Versioning¶

Suppose:

Policy v1
Policy v2
Policy v3

are all stored.

A query may retrieve all three unless filtering or lifecycle management is implemented.

A production system should define:

Active Version

and ensure obsolete versions do not incorrectly influence retrieval.

63. Recency-Aware Retrieval¶

Some applications care about freshness.

For example:

"What is the current travel reimbursement policy?"

The newest document should generally be preferred.

A retrieval architecture may combine:

Semantic Relevance
+
Metadata
+
Recency

This is an advanced ranking consideration.

64. Similarity Search Is Not Enough¶

A high similarity score does not guarantee:

Correct Answer

For example:

Query:
"What is the 2026 leave policy?"

Retrieved:
2023 Leave Policy

The content may be semantically similar but operationally wrong.

Therefore production retrieval requires:

Similarity
+
Metadata
+
Version
+
Security
+
Evaluation

65. Query Expansion¶

Some queries may be too short:

"leave"

A system may expand the query into:

employee annual leave entitlement
paid time off policy
leave carry-forward rules

These expanded queries can be searched independently and combined.

This is generally known as:

Multi-query retrieval / query expansion.

It will be explored in greater detail in the advanced retrieval module.

66. Similarity Search Failure Modes¶

Common failure cases include:

Wrong embedding model
Wrong similarity metric
Poor chunks
Poor query
Too-small K
Too-large K
Incorrect threshold
Missing metadata filters
Stale vectors
Duplicate vectors
Poor ANN configuration

Similarity search should therefore be evaluated as part of the complete retrieval pipeline.

67. Debugging Similarity Search¶

When retrieval fails:

User Query
    ↓
Inspect Query Embedding
    ↓
Inspect Top-K Scores
    ↓
Inspect Retrieved Chunks
    ↓
Inspect Metadata
    ↓
Inspect Index Configuration
    ↓
Inspect Embedding Model

Do not immediately assume the vector database is the problem.

68. Retrieval Debug Record¶

A useful diagnostic record could contain:

{
  "query_id": "q-102",
  "query": "What is the leave policy?",
  "top_k": 5,
  "results": [
    {
      "chunk_id": "policy-12",
      "score": 0.91
    },
    {
      "chunk_id": "policy-18",
      "score": 0.87
    }
  ],
  "latency_ms": 34
}

This supports retrieval debugging.

69. Observability¶

Track:

Query
Query ID
Embedding Model
Index
Top-K
Similarity Scores
Filters
Result IDs
Latency
Empty Result Rate

Be careful with logging sensitive query and document content.

Use appropriate data governance and privacy controls.

70. Production Similarity Search Workflow¶

1. Receive user query.

2. Validate query.

3. Resolve identity and tenant context.

4. Build authorization filters.

5. Generate query embedding.

6. Validate embedding dimension.

7. Select vector index.

8. Apply similarity search.

9. Apply security and metadata constraints.

10. Retrieve candidate results.

11. Apply similarity threshold if configured.

12. Apply optional ranking or reranking.

13. Remove duplicates.

14. Preserve source ordering where appropriate.

15. Return Top-K context.

16. Record retrieval metrics.

17. Pass approved context to the RAG pipeline.

71. Production Similarity Search Architecture¶

flowchart TD
    A["User Query"] --> B["Query Validation"]
    B --> C["Identity / Tenant Context"]
    C --> D["Authorization Filters"]

    B --> E["Query Embedding"]
    E --> F["Vector Index"]

    D --> F

    F --> G["ANN Candidate Search"]
    G --> H["Similarity Scoring"]

    H --> I["Metadata / Security Filtering"]
    I --> J["Threshold"]
    J --> K["Optional Reranking"]

    K --> L["Deduplication"]
    L --> M["Top-K Context"]

    M --> N["RAG Pipeline"]

72. Framework Example — LangChain¶

A simplified retrieval example:

retriever = vector_store.as_retriever(
    search_type="similarity",
    search_kwargs={
        "k": 5
    }
)

documents = retriever.invoke(
    "What is the annual leave policy?"
)

for document in documents:
    print(document.page_content)

The framework hides the lower-level vector database calls, but conceptually the workflow remains:

Query
 ↓
Embedding
 ↓
Vector Search
 ↓
Top-K

73. Framework Example — Similarity Threshold¶

A framework may expose threshold-based retrieval.

Conceptually:

retriever = vector_store.as_retriever(
    search_type="similarity_score_threshold",
    search_kwargs={
        "score_threshold": 0.75,
        "k": 10
    }
)

The actual supported behavior and score interpretation depend on the vector store and integration.

Always validate the returned scores.

74. Framework Example — LlamaIndex¶

A conceptual LlamaIndex retriever:

retriever = index.as_retriever(
    similarity_top_k=5
)

nodes = retriever.retrieve(
    "What is the annual leave policy?"
)

for node in nodes:
    print(node.text)

The underlying retrieval concepts remain independent of the framework.

75. Java Retrieval Interface¶

For a Java-first enterprise architecture:

public interface Retriever {

    List<RetrievedDocument> retrieve(
        Query query,
        RetrievalOptions options
    );
}

The retriever can use:

VectorStore
EmbeddingProvider
MetadataFilter
RankingStrategy

without exposing vendor-specific APIs to the application.

76. Retrieval Options¶

public record RetrievalOptions(
    int topK,
    Double similarityThreshold,
    SearchFilter filter
) {
}

This keeps retrieval configuration explicit.

77. Search Result Model¶

public record RetrievedDocument(
    String chunkId,
    String content,
    double score,
    Map<String, Object> metadata
) {
}

This provides a stable representation for downstream RAG components.

78. Retriever Architecture¶

flowchart LR
    A["RAG Application"] --> B["Retriever"]

    B --> C["EmbeddingProvider"]
    B --> D["VectorStore"]
    B --> E["SearchFilter"]
    B --> F["RankingStrategy"]

    D --> G["Vector Database"]

This keeps similarity search as a capability rather than a vendor-specific implementation detail.

79. Similarity Search Testing¶

Unit tests should cover:

Correct Top-K
Correct Score Ordering
Correct Filters
Empty Results
Threshold Behavior
Duplicate Results
Invalid Query
Dimension Mismatch

Integration tests should verify:

Embedding
+
Vector Store
+
Index
+
Metadata
+
Retrieval

80. Retrieval Regression Testing¶

When changing:

Embedding Model
Chunking
Index Type
ANN Parameters
Similarity Metric

run the retrieval evaluation suite again.

A seemingly small infrastructure change can alter retrieval quality.

81. Search Configuration as Code¶

Keep important retrieval configuration version-controlled.

Example:

retrieval:
  top-k: 10
  similarity-threshold: 0.72

  index:
    type: hnsw
    ef-search: 100

  filters:
    tenant-isolation: true

The exact configuration depends on the vector database.

82. Production Configuration Principles¶

Avoid hard-coding:

Top-K
Threshold
Index Parameters
Embedding Model

Instead make them:

Configurable
Versioned
Observable
Experimentable

83. Similarity Search Optimization¶

Optimization should follow measurement:

Measure
  ↓
Identify Bottleneck
  ↓
Change Configuration
  ↓
Benchmark
  ↓
Compare
  ↓
Deploy

Do not optimize only for latency while ignoring recall.

84. Retrieval Quality vs Latency¶

A production decision might look like:

Configuration	Recall@10	P95 Latency
ANN A	91%	20 ms
ANN B	95%	35 ms
ANN C	98%	90 ms

The correct choice depends on business requirements.

For a highly sensitive knowledge system, 98% recall may justify higher latency.

For an interactive application, 95% may be preferable.

85. Similarity Search Design Checklist¶

[ ] Embedding model selected

[ ] Query and document embeddings compatible

[ ] Vector dimension verified

[ ] Similarity metric selected

[ ] Exact vs ANN strategy selected

[ ] ANN index selected

[ ] ANN parameters configured

[ ] Top-K configured

[ ] Threshold evaluated

[ ] Metadata filtering implemented

[ ] Tenant isolation implemented

[ ] Security filtering implemented

[ ] Version filtering implemented

[ ] Empty retrieval handled

[ ] Duplicate results handled

[ ] Reranking strategy evaluated

[ ] Recall@K measured

[ ] Precision@K measured

[ ] MRR measured where appropriate

[ ] NDCG measured where appropriate

[ ] P50 latency measured

[ ] P95 latency measured

[ ] P99 latency measured

[ ] Throughput measured

[ ] Index build time measured

[ ] Memory measured

[ ] Storage measured

[ ] Retrieval logs implemented

[ ] Regression tests implemented

86. Common Mistakes¶

86.1 Using the Wrong Embedding Model¶

Query and documents must exist in a compatible vector space.

86.2 Choosing a Metric Arbitrarily¶

Similarity metrics should follow embedding model characteristics.

86.3 Always Using Exact Search¶

Exact search may become expensive at scale.

86.4 Blindly Using ANN Defaults¶

ANN parameters affect recall and latency.

86.5 Choosing an Arbitrary Threshold¶

Scores are model- and metric-dependent.

86.6 Using an Excessively Large K¶

More results can introduce noise and increase context cost.

86.7 Ignoring Security Filters¶

Similarity search must never bypass authorization.

86.8 Ignoring Document Versions¶

Old documents can outrank current policies.

86.9 Assuming Similarity Means Correctness¶

Similarity is a retrieval signal, not an answer guarantee.

86.10 Optimizing Only Latency¶

A fast retriever that misses relevant information is not useful.

87. Best Practices¶

1. Use compatible query and document embeddings.

2. Follow embedding-model guidance for similarity metrics.

3. Start with exact search for small datasets.

4. Move to ANN when scale requires it.

5. Benchmark ANN configurations.

6. Tune HNSW or IVF parameters using real queries.

7. Measure Recall@K.

8. Measure Precision@K.

9. Monitor P95 and P99 latency.

10. Combine semantic search with metadata filters.

11. Enforce tenant isolation.

12. Keep security filters outside prompt logic.

13. Handle stale documents.

14. Track document and embedding versions.

15. Calibrate similarity thresholds.

16. Evaluate Top-K experimentally.

17. Consider reranking for difficult retrieval workloads.

18. Use hybrid search when exact terms matter.

19. Log retrieval diagnostics safely.

20. Keep retrieval configuration version-controlled.

21. Separate candidate generation from ranking.

22. Treat vector search as one stage of the RAG pipeline.

23. Keep vector-store-specific APIs behind adapters.

24. Re-run retrieval evaluation after embedding or chunking changes.

25. Optimize quality and latency together.

88. Production Workflow¶

Source Documents
      ↓
Document Processing
      ↓
Chunking
      ↓
Embedding
      ↓
Vector Database
      ↓
ANN Index
      ↓
        ┌─────────────────────┐
        │                     │
User Query                Metadata
    ↓                        ↓
Query Embedding       Security Context
    ↓                        ↓
    └──────────┬─────────────┘
               ↓
        Similarity Search
               ↓
        Candidate Results
               ↓
        Filtering
               ↓
          Top-K Results
               ↓
          Optional Reranking
               ↓
        Context Assembly
               ↓
              LLM

89. Key Takeaways¶

Similarity search is the core operation behind semantic retrieval.
A user query is converted into an embedding before search.
Query and document embeddings must be compatible.
Common similarity/distance metrics include:
Cosine similarity
Dot product
Euclidean distance
Exact search compares the query against every stored vector.
ANN search reduces the search space for large datasets.
HNSW uses graph-based navigation.
IVF uses vector clustering.
Product Quantization can reduce memory usage.
ANN introduces a speed-versus-recall trade-off.
Top-K determines how many candidate results are returned.
Similarity thresholds can remove weak matches but must be calibrated.
Metadata filtering is essential for enterprise retrieval.
Tenant and security filtering must be enforced before unauthorized content reaches the LLM.
Hybrid search combines keyword and semantic retrieval.
Reranking can improve precision after candidate generation.
Recall@K measures whether relevant results are retrieved.
Precision@K measures the proportion of retrieved results that are relevant.
MRR evaluates the position of the first relevant result.
NDCG evaluates ranking quality across multiple relevance levels.
P95 and P99 latency are important production metrics.
Similarity scores do not equal answer confidence.
Vector search should be evaluated using representative queries and datasets.
Retrieval configuration should be version-controlled.
Changes to embeddings, chunking, metrics, or indexes should trigger regression evaluation.
Similarity search should remain an explicit capability in an enterprise architecture.

The central principle is:

Similarity search is not simply about finding the closest vectors; it is about finding the right evidence, under the right constraints, with the right balance between relevance, recall, latency, and cost.

Part IV — Prompt Engineering & RAG Fundamentals¶

Previous Chapter: 13. Vector Database Fundamentals

Current Chapter: 14 — Similarity Search Techniques

Next Chapter: 15. RAG Pipeline Components

Part IV Chapters¶

References¶

FAISS documentation
Chroma documentation
Qdrant documentation
pgvector documentation
Milvus documentation
Weaviate documentation
Pinecone documentation
Elasticsearch documentation
OpenSearch documentation
LangChain documentation
LlamaIndex documentation
Hugging Face documentation
Approximate Nearest Neighbor indexing documentation
Vector similarity search documentation
Enterprise search and retrieval architecture documentation

Enterprise AI Engineering Handbook
Building Production-Grade Enterprise AI Systems — One Chapter at a Time.