cevian

or try our brand-freaking-new free services:

```
curl -fsSL cli.tigerdata.com | sh
tiger auth login
tiger mcp install

```

How to Train Your Agent to Be a Postgres Expert | TigerData Turn AI into a Postgres expert with our MCP server. Get 35 years of best practices, versioned docs, and prompt templates for production-ready schemas.

We're releasing it today. Works with Claude Desktop, Cursor, Windsurf, VS Code. Open source with a free hosted service.

www.tigerdata.com/blog/free-po...

Instead of thin API wrappers, we teach AI things like "Always index foreign key columns - Postgres won't do it for you" and "Use BIGINT GENERATED ALWAYS AS IDENTITY, not SERIAL."

Last week @anthropic.com announced Skills - turning generalist LLMs into domain experts through structured knowledge. Different implementation, same philosophy. (Our approach might show how to implement Skills entirely through MCP 🤔)

That's where our get_prompt_template MCP tool comes in. Ask "design a schema for IoT devices" in plain language. The AI automatically pulls in the right Postgres expertise. No slash commands, no template picking.

The fix isn't obvious. I hate the copy-paste prompting I see in the wild. MCP prompt templates bother me too - why should users need slash commands? Isn't the AI supposed to figure this out?

AI coding should make this better. Instead, it's making things worse. Every "AI-generated" schema has the same issues: wrong ID types, missing indexes, timezone nightmares. Infuriating.

I've spent 10 years at Timescale (now TigerData) watching devs make the same database mistakes: forgetting to index foreign keys, case-sensitive email lookups, arbitrary VARCHAR limits. The pain of botched migrations and downtime is real.

For the past few months, I've been working on teaching AI coding tools to write better Postgres code. It's a passion project born from frustration.

github.com/timescale/pg...

The Database Has a New User—LLMs—and They Need a Different Database | TigerData TigerData experiments with self-describing Postgres using semantic catalogs. Early tests show 27% better SQL generation accuracy when AI agents understand schemas.

The results were strong, so we open sourced it.

www.tigerdata.com/blog/the-dat...

They just access the database and expect it to have the context.

But most databases can’t explain themselves. And that’s the core problem.

So we built a semantic catalog on top of Postgres. It added context inside the database. SQL accuracy improved by 27%.

I’ve been thinking about why LLMs are bad at text-to-sql.
Short answer: no context 🧵

Compared to developers, LLMs are more handicapped without context:
- They can’t read the application code.
- They can’t ask around.
- They don’t know where the docs are.

China is over 2 million with "administrative detentions", as I specified in my post.

Of course we also have a smaller prison population than China www.prisonstudies.org/country/china If you count "administrative detentions"

US prison population (biggest estimate, way more than actual): 2 million
world population: 8 billion
tell me how that's 5%?

Learn some math.

That's because Stalin killed *millions* of people he sent to the Gulag. You point is so stupid, it's painful.

Imagine not knowing Iran has a history of forcing gender reassignment on gay people

I wish people would analyze benchmarks with this kind of depth and insight more often.

Why can’t a gay man have a loving family you homophobe?

He is in no way a nazi but you are foul for making the word nazi meaningless in an age where real Nazis are gaining power

GitHub - timescale/pgai: A suite of tools to develop RAG, semantic search, and other AI applications more easily with PostgreSQL A suite of tools to develop RAG, semantic search, and other AI applications more easily with PostgreSQL - timescale/pgai

We take an index-like approach here github.com/timescale/pgai

🚫 Stop treating embedding generation like ETL. ✅ It’s indexing.

One of the most common misconceptions in the world of AI and LLMs: 👉 Vector embedding generation = ETL. Nope. It’s much more like building an index.

Conceptually, it's just like creating a full-text index on your data.

Pgai Vectorizer Now Works With Any Postgres Database By popular demand, pgai Vectorizer is now a Python library and CLI that works with any self-hosted or managed Postgres database. See how we built it.

🐘PGAI VECTORIZER NOW WORKS WITH ANY POSTGRES DATABASE (incl Timescale Cloud, Amazon RDS, Supabase, Azure PostgreSQL, and more)

www.timescale.com/blog/pgai-ve...

Best of all, my team is actively exploring this new direction. We think it’s the future. 🚀

#AI #PostgreSQL #DataIntegration #Innovation

With PostgreSQL, we have a unique opportunity:

A single platform that can power AI with access to both structured and unstructured data.

No silos. Just data.

Let’s skip the pain this time.

Let’s not arbitrarily separate our data when AI needs all of it—together.

💡 For AI applications, do we even need to follow that same path?

Why recreate the same divide between structured and unstructured data?

Listening to our CTO Michael Freedman talk about the evolution from data lakes (unstructured) and data warehouses (structured) into unified data lakehouses, I had a thought...

And check out the pgai GitHub here: github.com/timescale/pgai