Intelligent Search System

Keyword search fails when users describe what they want in natural language. “Good Italian place for a date night near me” returns nothing useful with BM25 alone because the match is semantic, not lexical. This project bridges that gap.

Architecture

The search pipeline has two retrieval stages followed by a ranking stage:

Stage 1 — Lexical Retrieval (BM25)

Standard Elasticsearch full-text search provides a fast first pass. This catches exact matches on restaurant names, cuisines, and locations. Results are returned in under 50ms.

Stage 2 — Semantic Retrieval (Vector Search)

User queries pass through a Hugging Face sentence-transformer model to produce dense embeddings. These embeddings are matched against pre-computed document embeddings stored in Elasticsearch’s dense vector index using cosine similarity.

LangChain orchestrates the query pipeline — handling query preprocessing, embedding generation, and result fusion from both retrieval stages.

Stage 3 — Cross-Encoder Reranking

Top candidates from both retrieval stages are reranked by a cross-encoder that jointly encodes the query and document. This is slower but more accurate, so it only runs on the top 20 candidates.

Vector Index Construction

Document embeddings are pre-computed offline using a sentence-transformer model fine-tuned on restaurant review data. Each document is encoded as a 384-dimensional vector and indexed in Elasticsearch with the dense_vector field type.

The index is rebuilt nightly to incorporate new restaurants and updated reviews. Incremental updates would be faster but introduce consistency issues — the nightly rebuild guarantees a uniform embedding space.

Query Processing

Natural language queries go through:

1. Intent classification: Is this a location query, a cuisine query, or a vibe query?
2. Entity extraction: Pull out cuisine types, price ranges, neighborhood names.
3. Query expansion: Augment the query with synonyms and related terms.

This preprocessing significantly improves retrieval recall for conversational queries.

Results

Metric	Before (BM25 only)	After (Hybrid)
Irrelevant results	Baseline	-40%
User engagement	Baseline	+20%
Daily queries	100K+	100K+
P95 latency	120ms	180ms

The 60ms latency increase from adding semantic retrieval was acceptable given the 40% reduction in irrelevant results. Users finding what they want faster more than compensates for the slight per-query slowdown.