# I Built a Hybrid Search Engine From Scratch — Here's What I Learned (LLM Zoomcamp 2026, Module 2)

> Source: <https://dev.to/derrickryangiggs/i-built-a-hybrid-search-engine-from-scratch-heres-what-i-learned-llm-zoomcamp-2026-module-2-3jdj>
> Published: 2026-06-26 11:49:10+00:00

I just completed Module 2 of the **LLM Zoomcamp 2026** by [@DataTalksClub](https://github.com/DataTalksClub/llm-zoomcamp/) — and this module completely changed how I think about search.

Module 1 taught me RAG and agentic pipelines. Module 2 taught me that the search step inside RAG matters far more than I realized — and that keyword search is only half the story.

Here's everything I built and learned.

Traditional keyword search matches words. If you search for "enroll", it finds documents containing "enroll" — but misses documents about "joining", "signing up", or "registration" even if they mean exactly the same thing.

**Vector search matches meaning, not words.**

Every piece of text gets converted into a vector — a list of hundreds of numbers that captures its semantic meaning. Similar meanings produce similar vectors, so you can find relevant documents even when they use completely different words.

This is the foundation of modern AI-powered search, and it's what makes RAG systems actually work at scale.

Instead of downloading the full PyTorch + CUDA stack (~2GB), I used a lightweight ONNX runtime embedder — same vectors, 30x smaller installation, runs on any CPU:

``` python
from embedder import Embedder

embedder = Embedder()  # loads Xenova/all-MiniLM-L6-v2 via ONNX
v = embedder.encode("How does approximate nearest neighbor search work?")
print(len(v))  # 384 dimensions
```

The model produces **384-dimensional vectors** — each number represents a dimension of meaning in the text.

Before using any library, I implemented vector search by hand to understand what's happening under the hood:

``` python
import numpy as np

# cosine similarity — vectors are normalized, so dot product works directly
def cosine_similarity(a, b):
    return np.dot(a, b)

# score all chunks against a query
scores = X.dot(v)  # X is the matrix of all chunk embeddings
best_idx = np.argmax(scores)
```

This is exactly what vector databases like Qdrant and pgvector do internally — just much faster at scale using HNSW indexing.

Full pages are too long and dilute the embedding — a match buried deep in a 10,000-character page still pulls in the whole page. The fix is chunking:

``` python
from gitsource import chunk_documents
chunks = chunk_documents(documents, size=2000, step=1000)
# 72 pages → 295 overlapping chunks
```

Overlapping chunks (step < size) ensure sentences at boundaries don't get cut off. After chunking, retrieval becomes far more precise.

`minsearch`

now has a `VectorSearch`

class that wraps the numpy math into a clean interface:

``` python
from minsearch import VectorSearch

vector_index = VectorSearch(keyword_fields=["filename"])
vector_index.fit(X, chunks)

results = vector_index.search(query_vector, num_results=5)
```

For the query **"How do I store vectors in PostgreSQL?"**:

`08-pgvector.md`

entirely because "pgvector" wasn't in the query`08-pgvector.md`

first because it understood the semantic connection between "store vectors" and "pgvector"This is the key insight: vector search finds meaning, keyword search finds words.

Neither approach is perfect on its own:

The solution is **hybrid search** — run both and merge the results using RRF:

``` python
def rrf(result_lists, k=60, num_results=5):
    scores = {}
    docs = {}
    for results in result_lists:
        for rank, doc in enumerate(results):
            key = (doc["filename"], doc["start"])
            scores[key] = scores.get(key, 0) + 1 / (k + rank)
            docs[key] = doc
    ranked = sorted(scores, key=scores.get, reverse=True)
    return [docs[key] for key in ranked[:num_results]]

results = rrf([vector_results, text_results])
```

RRF ignores raw scores (which live on different scales) and only looks at rank position. A document that ranks well in both lists beats one that's only strong in a single list — even if it wasn't first in either.

**1. Embeddings capture meaning, not words.** "Enroll" and "join" produce similar vectors. "Pizza" and "enrollment" don't. This is what makes semantic search powerful.

**2. Chunking is not optional.** Full pages dilute embeddings. 2,000-character overlapping chunks dramatically improve retrieval precision and cut LLM input tokens by 3x.

**3. Neither keyword nor vector search is best.** Use hybrid search (RRF) in production. It consistently outperforms either approach alone.

**4. ONNX makes embeddings practical anywhere.** No GPU, no PyTorch, no CUDA. 67MB download, runs on a basic laptop. There's no reason not to use vector search even in constrained environments.

**5. The right search approach depends on your data.** Vector search wins for semantic queries. Keyword search wins for exact terms (names, codes, IDs). Hybrid wins most of the time — but measure to be sure.

All my code for Module 2 is open source:

[github.com/Derrick-Ryan-Giggs/llm-zoomcamp-2026](https://github.com/Derrick-Ryan-Giggs/llm-zoomcamp-2026)

It includes:

`vector-search.ipynb`

— embeddings, Qdrant, and vector RAG pipeline`Vector Search Homework.ipynb`

LLM Zoomcamp is **completely free** — no paywall, no certificate fees.

Sign up: [github.com/DataTalksClub/llm-zoomcamp](https://github.com/DataTalksClub/llm-zoomcamp/)

*Are you working through LLM Zoomcamp 2026? Drop a comment — I'd love to compare notes.*