Manik khandelwal

Engineer @ Microsoft | Tweets about Science, Engineering, Nature, Travel, Entrepreneurship and Philosophy

[Manik khandelwal @Manikkhandelwa2]

Feels like there's still room for a true capability-driven database SDK—one that standardizes common operations while letting databases expose their strengths through optional capabilities.

Would love to know if anyone is building something along these lines.

#Databases #SDK

[Manik khandelwal @Manikkhandelwa2]

Why does every #database require learning a completely different #SDK?

Not because CRUD is different—but because every database exposes its own unique capabilities (#vector #search, change streams, graph queries, #full-text search, etc.).

🧵

[Manik khandelwal @Manikkhandelwa2]

We've already seen this pattern work in other ecosystems:

• LangChain → VectorStore interface
• Apache Calcite → Adapter architecture
• Trino → Connector model
• Prisma → Driver adapters
Each abstracts a different layer—but the idea is remarkably similar.

[Manik khandelwal @Manikkhandelwa2]

@samhogan Wow, are these numbers real? and how about the accuracy and tuning experiece?

[Manik khandelwal @Manikkhandelwa2]

@prakdadlani Really nice to have

[Manik khandelwal @Manikkhandelwa2]

My takeaway:

The future isn't just better prompts.

It's better context.

Choosing what to send may become just as important as choosing which model to use.

Curious—if you were building an AI agent, where do you think most token waste comes from?

Logs? JSON? RAG? Chat history?

[Manik khandelwal @Manikkhandelwa2]

If an LLM has a 1M-token context window...

...does that mean you should send it 1M tokens? 🤔

I don't think so.

Here's why 👇

[Manik khandelwal @Manikkhandelwa2]

One thing people often mix up:

Semantic compression ≠ Prompt caching.

Compression reduces the number of tokens.

Caching avoids recomputing the same prompt prefix.

Different techniques.

Same goal: lower cost and faster inference.

[Manik khandelwal @Manikkhandelwa2]

@AyushiChadha24 Interested in your approach—how are you managing without peers or labs?

[Manik khandelwal @Manikkhandelwa2]

@DrJimFan @PRERITKHANDELW3

[Manik khandelwal @Manikkhandelwa2]

@christinaingoog interested, please dm.