ceviantech.bsky.social
Tech lead for pgai, pgvectorscale, and all other AI things @ Timescale.
Postgres rocks!
58 posts
61 followers
136 following
Regular Contributor
Active Commenter
comment in response to
post
Why canโt a gay man have a loving family you homophobe?
comment in response to
post
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
comment in response to
post
We take an index-like approach here github.com/timescale/pgai
comment in response to
post
Best of all, my team is actively exploring this new direction. We think itโs the future. ๐
#AI #PostgreSQL #DataIntegration #Innovation
comment in response to
post
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.
comment in response to
post
Letโs skip the pain this time.
Letโs not arbitrarily separate our data when AI needs all of itโtogether.
comment in response to
post
๐ก For AI applications, do we even need to follow that same path?
Why recreate the same divide between structured and unstructured data?
comment in response to
post
And check out the pgai GitHub here: github.com/timescale/pgai
comment in response to
post
๐ช๐ต๐ ๐ฑ๐ถ๐ฑ ๐๐ฒ ๐ฏ๐๐ถ๐น๐ฑ ๐๐ต๐ถ๐?
At Timescale we believe Postgres is the ideal home for both your structured data and vector embeddings. By keeping embeddings automatically synchronized to source documents in S3, Vectorizer ensures your Postgres database remains the single source of truth for your AI systems.
comment in response to
post
๐ฉ๐ฒ๐ฐ๐๐ผ๐ฟ๐ถ๐๐ฒ๐ฟ ๐๐๐ฝ๐ฝ๐ผ๐ฟ๐๐ ๐ฎ ๐๐ถ๐ฑ๐ฒ ๐๐ฎ๐ฟ๐ถ๐ฒ๐๐ ๐ผ๐ณ ๐ณ๐ถ๐น๐ฒ ๐ณ๐ผ๐ฟ๐บ๐ฎ๐๐ including PDF, DOCX, TXT, XLSX, PPTX, images, HTML, and more.
comment in response to
post
๐ฝ๐ด๐ฎ๐ถ ๐ฉ๐ฒ๐ฐ๐๐ผ๐ฟ๐ถ๐๐ฒ๐ฟ ๐ฒ๐ป๐ฎ๐ฏ๐น๐ฒ๐ ๐ฑ๐ฒ๐๐ฒ๐น๐ผ๐ฝ๐ฒ๐ฟ๐ ๐๐ผ:
โ
Spend less time wrangling data infrastructure with automatic updating and synchronization.
โ
Continuously improve your AI systems by testing different embedding models or chunking strategies with a single line of SQL.
comment in response to
post
Pgai Vectorizer now provides a streamlined approach where you can reference documents in S3 via URLs stored in a database table. The vectorizer handles the complete workflowโdownloading documents, parsing them to extract content, chunking text, and generating embeddings for vector search.
comment in response to
post
This complexity creates a brittle infrastructure that inevitably leads to stale embeddings and wasted engineering hours.
comment in response to
post
The problem: Before pgai Vectorizer, developers building RAG applications had to manage complex ETL pipelines, multiple systems, data synchronization services, queuing systems, and monitoring tools just to keep document embeddings up-to-date.
comment in response to
post
Full Pgvector vs Qdrant benchmark blog post:
www.timescale.com/blog/pgvecto...
comment in response to
post
github.com/timescale/pg...
comment in response to
post
๐งโ๐ป Sounds exciting! How can I get started?
Pgvectorscale is open-source under the PostgreSQL license, and free to use on any PostgreSQL database. Itโs available on any database service on Timescale Cloud, and you can find installation instructions on the pgvectorscale GitHub repository.
comment in response to
post
While there are niche use cases that do benefit from a dedicated vector database like Qdrant, pgvector and pgvectorscale enables 99% of developers to start and scale confidently with just PostgreSQL.
comment in response to
post
๐ค Why does this matter?
Postgres is the database that many developers already know, use, and trust. And this benchmark shows that it can not only keep up, but even outperform specialized vector databases on high-scale workloads
comment in response to
post
Qdrant does deliver better query latencies (between 39-48% better than Postgres), showing less variance between percentiles, which is to be expected from a purpose built vectordb.
comment in response to
post
๐ How did Postgres perform vs Qdrant?
Postgres with pgvector and pgvectorscale outperforms Qdrant on throughput by 11.4x and delivers sub-75ms p99 query latencies at 99% recall.
comment in response to
post
And we put it to the test against Qdrant, the leading specialized vector database, on an ANN benchmark of 50M embeddings. The results surprised even us...
comment in response to
post
๐There's a common misconception that Postgres and pgvector can't scale for vectors. Thatโs why we
(TimescaleDB) built pgvectorscale, an open-source PostgreSQL extension that supercharges pgvector with greater performance for large scale vector workloads.
comment in response to
post
This essay is so mid
comment in response to
post
Yes. Great result. We were surprised to see it!
comment in response to
post
Read more about the new benchmark here: www.timescale.com/blog/benchma...
comment in response to
post
I should say, the 3rd is especially exciting to me because no major benchmark even really tries to address materialized data. And yet, pretty much every application that has a dashboard over historical data does this! It's a weird blind spot in the industry.
comment in response to
post
We identified 3 major changes:
1. Queries join multiple tables
2. Queries filter on specific objects and time windows
3. Pre-computed aggregations ensure instant responses
comment in response to
post
check out the HF docs for more info huggingface.co/docs/dataset...
comment in response to
post
Ha! That's the exact same explanation I use. I also think that this visualization helps with a lot of other concepts. But it's never taught in school (at least in the U.S.)
comment in response to
post
Can I please get added? Thanks!
comment in response to
post
Any chance I am worthy of joining this highly distinguished list?
comment in response to
post
Collab notebook: colab.research.google.com/drive/1rugFO...