Hey everyone,

A few days back I shared my experiments with hybrid search (combining traditional lexical search with vector/semantic search). Well, I've been busy, and I'm back with some **major upgrades** that I think you'll find interesting.

**TL;DR:** We now have 1024-dimensional embeddings, blazing fast GPU inference, and you can generate embeddings via our free API endpoint. Plus: you can literally search with emojis now. Yes, really. 🚲 finds bicycles. 🐕 finds dog jewelry. Keep reading.

---

## What Changed?

### 1. Upgraded from 384D to 1024D Embeddings

We switched from `paraphrase-multilingual-MiniLM-L12-v2` (384 dimensions) to `BAAI/bge-m3` (1024 dimensions).

**Why does this matter?**

Think of dimensions like pixels in an image. A 384-pixel image is blurry. A 1024-pixel image is crisp. More dimensions = the model can capture more nuance and meaning from your text.

The practical result? Searches that "kind of worked" before now work **really well**, especially for:

- Non-English languages (Romanian, German, French, etc.)

- Domain-specific terminology

- Conceptual/semantic queries

### 2. Moved Embeddings to GPU

Before: CPU embeddings taking 50-100ms per query. Now: GPU embeddings taking ~2-5ms per query.

The embedding is so fast now that even with a network round-trip from Europe to USA and back, it's **still faster** than local CPU embedding was. Let that sink in.

### 3. Optimized the Hybrid Formula

After a lot of trial and error, we settled on this normalization approach:

```

score = vector_score + (lexical_score / (lexical_score + k))

```

Where `k` is a tuning parameter (we use k=10). This gives you:

- Lexical score normalized to 0-1 range

- Vector and lexical scores that play nice together

- No division by zero issues

- Intuitive tuning (k = the score at which you get 0.5)

### 4. Quality Filter with frange

Here's a pro tip: use Solr's `frange` to filter out garbage vector matches:

```

fq={!frange l=0.3}query($vectorQuery)

```

This says "only show me documents where the vector similarity is at least 0.3". Anything below that is typically noise anyway. This keeps your results clean and your users happy.

---

## Live Demos (Try These!)

I've set up several demo indexes. **Each one has a Debug button in the bottom-right corner** - click it to see the exact Solr query parameters and full `debugQuery` analysis. Great for learning!

### 🛠️ Romanian Hardware Store (Dedeman)

Search a Romanian e-commerce site with emojis:

[🚲 → Bicycle accessories](https://opensolr.com/search/dedeman?topbar=block&q=%F0%9F%9A%B2&in=web&og=yes&locale=&duration=&source=&fresh=no&lang=)

No keywords. Just an emoji. And it finds bicycle mirrors, phone holders for bikes, etc. The vector model understands that 🚲 = bicicletă = bicycle-related products.

### 💎 English Jewelry Store (Rueb.co.uk)

Sterling silver, gold, gemstones - searched semantically:

[🐕 → Dog-themed jewelry](https://opensolr.com/search/rueb?topbar=block&q=%F0%9F%90%95&in=web&og=yes&locale=&duration=&source=&fresh=no&lang=)

[⭐️ → Star-themed jewelry](https://opensolr.com/search/rueb?topbar=block&q=%E2%AD%90%EF%B8%8F&in=web&og=yes&locale=&duration=&source=&fresh=no&lang=)

### 🧣 Luxury Cashmere Accessories (Peilishop)

Hats, scarves, ponchos:

[winter hat → Beanies, caps, cold weather gear](https://opensolr.com/search/peilishop?topbar=block&q=winter+hat&in=web&og=yes&locale=&duration=&source=&fresh=no&lang=)

### 📰 Fresh News Index

Real-time crawled news, searchable semantically:

[🍳 → Food/cooking articles](https://opensolr.com/search/vector?topbar=block&q=%F0%9F%8D%B3&in=web&og=yes&locale=&duration=&source=&fresh=no&lang=)

[what do we have to eat to boost health? → Nutrition articles](https://opensolr.com/search/vector?topbar=block&q=what+do+we+have+to+eat+to+boost+health%3F&in=web&og=yes&locale=&duration=&source=&fresh=no&lang=)

This last one is pure semantic search - there's no keyword "boost" or "health" necessarily in the results, but the *meaning* matches.

---

## Free API Endpoint for 1024D Embeddings

Want to try this in your own Solr setup? We're exposing our embedding endpoint for free:

```bash

curl -X POST https://opensolr.com/api/embed \

-H "Content-Type: application/json" \

-d '{"text": "your text here"}'

```

Returns a 1024-dimensional vector ready to index in Solr.

**Schema setup:**

```xml

<fieldType name="knn_vector" class="solr.DenseVectorField"

vectorDimension="1024" similarityFunction="cosine"/>

<field name="embeddings" type="knn_vector" indexed="true" stored="false"/>

```

---

## Key Learnings

1. **Title repetition trick**: For smaller embedding models, repeat the title 3x in your embedding text. This focuses the model's limited capacity on the most important content. Game changer for product search.

2. **topK isn't "how many results"**: It's "how many documents the vector search considers". The rest get score=0 for the vector component. Keep it reasonable (100-500) to avoid noise.

3. **Lexical search is still king for keywords**: Hybrid means vector helps when lexical fails (emojis, conceptual queries), and lexical helps when you need exact matches. Best of both worlds.

4. **Use synonyms for domain-specific gaps**: Even the best embedding model doesn't know that "autofiletantă" (Romanian) = "drill". A simple synonym file fixes what AI can't.

5. **Quality > Quantity**: Better to return 10 excellent results than 100 mediocre ones. Use `frange` and reasonable `topK` values.

---

## What's Next?

Still exploring:

- Fine-tuning embedding models for specific domains

- RRF (Reciprocal Rank Fusion) as an alternative to score-based hybrid

- More aggressive caching strategies

Happy to answer questions. And seriously, click that Debug button on the demos - seeing the actual Solr queries

is super educational!

---

*Running Apache Solr 9.x on [OpenSolr.com](https://opensolr.com/) - free hosted Solr with vector search support.*