The Power of Hybrid Search: Combining Vector and Full-text Search

Introduction

In the era of big data, information retrieval goes beyond simple keyword matching—it requires understanding semantics and context. Prisma AI uses Hybrid Search technology to ensure optimal accuracy when retrieving information from personal and enterprise knowledge bases.

1. Two Pillars of Hybrid Search in Prisma AI

Prisma AI's search system is built on the combination of two complementary methods:

Vector Search (Semantic Search)

Uses embedding models to transform text into mathematical vectors. This method allows the system to find information based on meaning and concepts, even when users don't use the exact words found in the documents.

Technical Features:

• Uses HNSW (Hierarchical Navigable Small World) indexes
• Applies cosine distance for vector queries
• Extremely fast query speed with high accuracy

Full-text Search (Keyword Search)

Leverages PostgreSQL's capabilities with tsvector and tsquery to perform traditional keyword searches. This method is extremely effective for finding:

• Technical terminology
• Error codes
• Specific names

These are cases where vector search might sometimes miss. The system utilizes GIN indexes to optimize text lookups.

2. Reciprocal Rank Fusion (RRF) Algorithm: The Heart of Accuracy

The breakthrough of Prisma AI lies in using the Reciprocal Rank Fusion (RRF) algorithm to merge results from both search methods.

RRF Processing Workflow

Step 1: Separate Ranking

The system performs Semantic Search and Keyword Search queries in parallel, then assigns each document a rank based on relevance in each method.

Step 2: Calculate RRF Score

Prisma AI applies the formula:

Score = 1.0 / (k + Rank_Vector) + 1.0 / (k + Rank_Text)

Where the constant k is typically set to 60 to balance weights between the two methods.

Step 3: Merge Results

Documents from both sources are combined through a FULL OUTER JOIN. Documents appearing in high positions in both ranking lists receive the highest combined scores.

Step 4: Sort and Filter

Finally, the system re-sorts the entire list by descending RRF score and extracts only the most relevant chunks to send to the LLM for processing.

RRF Process Illustration

┌─────────────────┐     ┌─────────────────┐
│  Vector Search  │     │ Full-text Search│
│    (Semantic)   │     │   (Keyword)     │
└────────┬────────┘     └────────┬────────┘
         │                       │
         ▼                       ▼
┌─────────────────┐     ┌─────────────────┐
│   Rank: 1,2,3   │     │   Rank: 1,2,3   │
│   Doc A, B, C   │     │   Doc B, D, A   │
└────────┬────────┘     └────────┬────────┘
         │                       │
         └───────────┬───────────┘
                     ▼
         ┌───────────────────────┐
         │   FULL OUTER JOIN     │
         │   + RRF Score Calc    │
         └───────────┬───────────┘
                     ▼
         ┌───────────────────────┐
         │  Final Ranked List    │
         │  Doc B > A > C > D    │
         └───────────────────────┘

3. Why is Hybrid Search Important?

This combination helps Prisma AI overcome the limitations of each individual method:

Increased Recall

Ensures no information is missed thanks to the ability to search by both keywords and meaning. When one method misses, the other compensates.

Increased Precision

The RRF algorithm prioritizes truly relevant documents, helping AI deliver focused answers and avoid hallucination.

Transparency with Citations

Thanks to Hybrid Search finding the exact original text passages, Prisma AI can attach accurate Citations to specific document IDs, allowing users to easily verify sources.

Conclusion

With powerful Hybrid Search technology, Prisma AI is not just a search tool, but a digital brain capable of finding and connecting the smallest pieces of knowledge in your massive data repository in a scientific and reliable manner.

The combination of Vector Search and Full-text Search, orchestrated by the RRF algorithm, creates a superior search system—where semantics and keywords work together to deliver the best results for users.

Want to experience the power of Hybrid Search in Prisma AI? Contact us for consultation and product demo.