Hybrid search (semantic + keyword) in Prompt Engineering / GenAI - Model Pipeline Trace

Model Pipeline - Hybrid search (semantic + keyword)

This hybrid search pipeline combines keyword matching with semantic understanding to find the most relevant documents. It first filters documents by keywords, then ranks them by semantic similarity using a trained model.

Data Flow - 5 Stages

1Input Query

1 query string→User inputs a search query→1 query string

"best Italian restaurants near me"

↓

2Keyword Filtering

N documents x text→Filter documents containing query keywords→M documents x text (M ≤ N)

From 1000 docs, filter to 150 docs containing words like 'Italian', 'restaurants'

↓

3Semantic Embedding

1 query string and M documents x text→Convert query and documents to vector embeddings→1 query vector (768 dims), M document vectors (768 dims each)

Query vector: [0.12, -0.05, ..., 0.33], Document vector: [0.10, -0.02, ..., 0.30]

↓

4Similarity Scoring

1 query vector, M document vectors→Calculate cosine similarity between query and each document vector→M similarity scores (float between -1 and 1)

[0.85, 0.78, 0.65, ...]

↓

5Ranking and Output

M documents with similarity scores→Sort documents by similarity score descending→Top K documents ranked

Top 5 documents with scores: [(doc23, 0.85), (doc7, 0.83), ...]

Training Trace - Epoch by Epoch

Loss
0.7 |****
0.6 |*** 
0.5 |**  
0.4 |*   
0.3 |*   
0.2 |    
     1 2 3 4 5 Epochs

Epoch	Loss ↓	Accuracy ↑	Observation
1	0.65	0.60	Model starts learning semantic relations, initial moderate accuracy
2	0.48	0.72	Loss decreases, accuracy improves as embeddings better capture meaning
3	0.35	0.81	Model converges, semantic similarity scores become more reliable
4	0.30	0.85	Fine tuning improves ranking quality, loss stabilizes
5	0.28	0.87	Final epoch shows best balance of loss and accuracy

Prediction Trace - 4 Layers

Layer 1: Keyword Filtering

Layer 2: Semantic Embedding

Layer 3: Similarity Scoring

Layer 4: Ranking

Model Quiz - 3 Questions

What is the main purpose of the keyword filtering stage?

A To reduce the number of documents before semantic comparison

B To convert text into vectors

C To calculate similarity scores

D To rank documents by relevance

Key Insight

Combining keyword filtering with semantic similarity balances speed and understanding, enabling efficient and meaningful search results.