This repository presents a minimal and effective implementation of semantic document retrieval using vector embeddings and cosine similarity. The system is designed to identify the document that best matches the meaning of a user's input query.
Traditional search methods rely on keyword overlap. This method improves upon that by leveraging pretrained transformer-based sentence embeddings, allowing retrieval based on semantic similarity rather than lexical similarity.
The process involves:
- Embedding each document and query into a high-dimensional space.
- Computing cosine similarity between the query vector and document vectors.
- Returning the document with the highest semantic alignment.
Each document d_i and query q is passed through a transformer model f, which encodes the text into an embedding vector:
d_i = f(d_i), q = f(q)
Where f maps raw text to ℝ^384 using a pre-trained model.
Similarity between query and document embeddings is calculated as:
cos_sim(q, d_i) = (q · d_i) / (||q|| * ||d_i||)
This metric reflects the angle between vectors. Higher values indicate greater similarity in meaning.
All cosine scores are computed, sorted, and the top document is returned:
sorted_scores = sorted(enumerate(scores), key=lambda x: x[1])
document_index, score = sorted_scores[-1]- Model Name:
sentence-transformers/all-MiniLM-L6-v2 - Architecture: Distilled bi-encoder (Transformer-based)
- Embedding Dimension: 384
- Provider: HuggingFace Model Hub
query = "Kylie"Returns:
query: Kylie
best_match: Kylie Jenner commands attention with her signature style, glamorous beauty, and trendsetting influence on social media.
similarity_score: 0.7823
Install required libraries:
pip install -r requirements.txtpython /Users/manavdesai/Desktop/GitHub/Langchain_Models/document_search_app.py.py