$ whoami

  • Kumar Shivendu

  • Software Engineer @ Qdrant

  • I ❤️ information retrieval, performance, and data mining.

  • First talk!

  • Qdrant: Shaping the future of search and beyond

  • Approaches:

    • Keyword match, Regex, Boolean operators
    • Extracting metadata using NLP and CV
    • Knowledge graphs, Vocabs

  • Challenges:

    • Effort to maintain the knowledge
    • Growth of unstructured data
    • Multimodal search remains hard: Text, Image, Audio, Video

Vectors

  • Points in an N-dimensional space
  • Anything -> Vector
  • Generated from:
    • ML models
    • Metric learning
      • CLIP
  • Nearest points
  • Example: Google Lens
  • But this is expensive and not easy to scale
  • Solution: Indexing and approximation

What is Qdrant

  • Vector Search Engine (aka Vector DB)

  • 15k+ stars on Github

  • Written in Rust 🦀

  • SDKs for Python, JS, Go, Java, etc

  • Twitter, Canva, Meesho, Flipkart

The HNSW Index

  • Skip Lists + Graphs
  • Approximate and Tunable
  • Filter during search
  • Quantization
POST /collections/rentals/points/search
{
  "query": [0.2, 0.3, 0.4, 0.5], // vector generated from image/text/video
  "filter": { "must": [{"key": "locality", "match": {"value": "Indiranagar"}}] },
  "limit": 10
}

  • [
  {"id": 4, "score": 0.56, "payload": {...}},
  {"id": 2, "score": 0.40, "payload": {...}},
  {"id": 5, "score": 0.23, "payload": {...}},
]

Beyond search: Recommendations

  • Realtime addition of points is possible.
  • average_vector and best_score
  • POST /collections/rentals/points/recommend
{
  "positive": [100, 231], // vector ids
  "negative": [718, [0.2, 0.3, 0.4, 0.5]], // vector id and vector
  "filter": { "must": [{"key": "locality", "match": {"value": "Indiranagar"}}] },
  "strategy": "best_score",
}
  • DailyMotion (Qdrant), Spotify (Annoy)

How to find this thing on the internet?

  • No reverse image search
  • No known name

Strategy One

  • Describe the thing
    • "Combination of human, dragon and chicken"
    • "Mythology creature of human and dragon"

Strategy Two

  • Search for similar images
    • Similarity bubble

Beyond search: Discovery API

  • Unique iterative search by Qdrant

  • Combine multi-modal vectors in single query

  • POST collections/my-collection/points/discover
{
  "target": [0.63, 0.10, 0.91, 0.55],
  "context": [
    {
      "positive": 7125, // <-- ID of the example
      "negative": 122   // <-- This can also be a vector
    }
  ],
}

Summary

  • Anything => vector

  • Vectors >> similarity search

  • Thousands of use-cases with Qdrant

  • Find me at
    • kshivendu.dev/bio
    • kshivendu.dev/twitter

--- ## Topics to cover * Evolution of search * Traditional approaches and drawbacks * Vectors search and how it harnesses ML models * Qdrant * Building HNSW index and vector search * Beyond similarity search: * Recommendations * Discovery * Sparse vectors

Replace search space image

Image showing vector search

--- ## How discovery uses that? ##### Remember Metric Learning? ![bg 80%](./imgs/triplet-loss.png) ---

![bg 95%](./imgs/discovery-context.png) ![bg 90%](./imgs/context-pairs.png) ![bg 90%](./imgs/context-with-target.png) ---

## How multi-modal embeddings look like? ![bg 90%](./imgs/cross-modal-space.png) --- ![bg](./imgs/clip-discovery.png) --- ![bg](./imgs/complex-context-search.png) ---

## Sparse vectors * VS text search * BM25 & TF-IDF * Transformer's attention weights * SPLADE ![bg right:60% 90%](./imgs/sparse-vectors.png) ---

* Navigating search (read vector) space is powerful!

* Thank you!