AI-Infused Applications: The Rise of Vector and Graph Databases

Vector Embeddings

cat=[1.5,−0.4,7.2,19.6,3.1]

How vector embeddings are generated?

Searching/Querying in a vector DB

Relevance

Piecing it all together

Answer the following user query by only using the information provided below: 

- {{ text chunks }}

User query is: {{ query }}

Knowledge graphs

At this point you can jump into creating a RAG pipeline right away and the LLM would have adequate information to answer straight factual questions.

Use cases for RAG

• Chat bot answering questions about refund policy
• Customer support for product information
• Simple FAQ systems

However a mere similarity search would not yield some of the intricate details that may be crucial in providing an accurate answer.

Consider following story:

• John was happily married to Katie for ten years
• Katie is a childhood friend of Martha
• Martha is Jake’s wife
• Jake and Trevor are best friends
• Trevor had a fight with John
• After Jake learned about this fight, he started hating Trevor
  

How a Graph Database Helps

Building a knowledge graph

Graph retrieval

{
  people_search(func: eq(entity, "John Doe")) {
    entity
    entity_type
    relation
    object
    object_type
  }
}

{
  "data": {
    "people_search": [
      {
        "entity": "John Doe",
        "entity_type": "Person",
        "relation": "OCCUPATION",
        "object": "Businessman",
        "object_type": "Occupation"
      },
      {
        "entity": "John Doe",
        "entity_type": "Person",
        "relation": "DEATH_CAUSE",
        "object": "Murder",
        "object_type": "CauseOfDeath"
      }
    ]
  }
}

Conclusion

In this guide, we’ve explored the world of vector databases and their applications in text retrieval. We’ve seen how vector embeddings can be used to search for relevant text segments, and how the results can be further augmented with contextual information using Large Language Models (LLMs).

We’ve also delved into the realm of graph databases, which excel at capturing complex relationships between entities. By modeling these relationships as nodes and edges, we can create a powerful knowledge graph that can help us answer nuanced questions.

The key takeaways from this guide are:

• Vector databases offer a powerful way to search for relevant text segments
• LLMs can be used to augment the context with additional information
• Graph databases provide a powerful way to capture complex relationships between entities
• Knowledge graphs can help us answer nuanced questions by modeling these relationships as nodes and edges

To build a comprehensive knowledge graph, we’ll need to:

• Extract key entities and relationships from user prompts using NLP models
• Model these relationships as nodes and edges in a graph database
• Add full-text indices on entity names for efficient text search
• Use LLMs to augment the context with additional information

Content under the appendix is only meant for my reference for future post. If you find my guide useful please do share it :)

Appendix

Graph databases

• Memgraph
• Kùzu

Qdrant - a vector database

Qdrant is a relatively new addition to vector database space. What I like the most about Qdrant is how easy it is to get started development using it. You just download the platform binary here and launch 🚀

ulimit -n 10000
./qdrant

I had to set the ulimit to avoid getting error: Too many open files 👀

Getting the frontend working

Although in official docs say that Qdrant’s web ui client can be accessed through http://localhost:6333/dashboard it doesn’t work out of the box at the time of writing. You have to clone their web gui repo and run npm:

git clone https://github.com/qdrant/qdrant-web-ui
cd qdrant-web-ui
npm i
npm start

Now you should be able to access it through http://localhost:5173 (you need to run qdrant in a separate terminal :duh 🙄 )