With this single video - I was able to understand what I need to become a AI Engineer by leveraging PostgresSQL extensions and vector databases and then to consume them in my React/React Native Applications. Thanks a ton for making these videos. The instructor needs a raise for putting everything so well together.
Hi, your presentation gives me a clear idea of vector DB (I am new to it). May I ask a question about the example in 18:03 If I understand whats going on correctly, you are encoding every row's columns (or their combinations) to vector data type, and then the same to the verbose text query using llm model text-embedding-3-small, so you can compare them with cosine similarity and output the top result, therefore we can get data in the columns and send them WITH the original verbose text query to the llm as usual. 1. Is this the idea of what RAG does? 2. If so, since what is being stored in the row/columns are raw data (string/int whatever), for example a date could be a expiry date of a password/member since/birthday/etc anything, should we embedded the original data directly, or turn them verbose first BEFORE encoding to vector, in order to get a better result? Or depends on the situations? For example sex, age, verbose description, embedded(verbose description), embedded([sex, age]) m, 18, "this is a man in age 18", [0.01, .......], [.........val in vector]
lovely like usual , there is indeed allot to learn but were getting closer :) Bless you, ps. Regarding storing structured and unstructured data in the same table : are you using the technique to store complete structured tables inside a JSONB ? We thought about this approach but dropped it in favor of separating structured from unstructured data to prevent mis-matching and allow for better isolation/scaling. Still experimenting but currently our set-up creates 1 table per structured document and infer the SCHEMA dynamically on upload + the embedding. Than the Agent decides on run-time which tables to query on. Unstructured documents can be bundled together more easily but placing all document types together can give false-positive search results ?
Mmm... Old school engineer here, spent more 30 years with db systems. And now I understand I don't want a black box RAG system but I want to implement AI stuff with PG ! For me there is a little thing you can do better: add some examples for retrieving data without and with the vectors, especially when there is a where clause. Other than that, your video is a big source of inspiration. Thanks a lot.
Well, after trying the whole thing I think the caveat here is to use pgai that depends on OpenAI: this is not local, you have to pay for the tokens, your data are going away and it is a black box. So I found another way, coding some functions locally in Postgres to use ollama locally with local models: no privacy leak or data leak, no token cost. This is what I understood but I am a rookie.
pgai supports Ollama so you can use local models for greater privacy and lower cost. Check out the Github repo for more. The example used in the video is with OpenAI, but pgai also supports Ollama, Cohere, and Anthropic models too.
Also, where is StreamingDiskANN come from? Seems only IVFFLAT and HNSW are here but diskann SQL Error [42704]: ERROR: access method "diskann" does not exist! Ha, got it: pgvectorscale !
Appreciate the video. Just FYI, you're pronouncing the word "build" like "bulled" throughout the video, but most native speakers pronounce it like "billed"
Excellent, excellent, excellent. It does even more than I had merely hoped for. Forgive me, postgres, I should never have doubted you.
Massivelly underrated video! So much valuable info, thank you so much!
Thanks! Glad you found it helpful.
With this single video - I was able to understand what I need to become a AI Engineer by leveraging PostgresSQL extensions and vector databases and then to consume them in my React/React Native Applications.
Thanks a ton for making these videos.
The instructor needs a raise for putting everything so well together.
That's awesome - thanks for sharing! Glad we could help.
Amazing video, learnt a lot. Will make time to read timescale blog regularly.
Thanks! much appreciated
You should do more tutorials like these
What additional topics would you like to see? Let us know and we can make it happen.
Hi, your presentation gives me a clear idea of vector DB (I am new to it). May I ask a question about the example in 18:03
If I understand whats going on correctly, you are encoding every row's columns (or their combinations) to vector data type, and then the same to the verbose text query using llm model text-embedding-3-small, so you can compare them with cosine similarity and output the top result, therefore we can get data in the columns and send them WITH the original verbose text query to the llm as usual.
1. Is this the idea of what RAG does?
2. If so, since what is being stored in the row/columns are raw data (string/int whatever), for example a date could be a expiry date of a password/member since/birthday/etc anything, should we embedded the original data directly, or turn them verbose first BEFORE encoding to vector, in order to get a better result? Or depends on the situations? For example
sex, age, verbose description, embedded(verbose description), embedded([sex, age])
m, 18, "this is a man in age 18", [0.01, .......], [.........val in vector]
lovely like usual , there is indeed allot to learn but were getting closer :) Bless you,
ps. Regarding storing structured and unstructured data in the same table : are you using the technique to store complete structured tables inside a JSONB ?
We thought about this approach but dropped it in favor of separating structured from unstructured data to prevent mis-matching and allow for better isolation/scaling.
Still experimenting but currently our set-up creates 1 table per structured document and infer the SCHEMA dynamically on upload + the embedding. Than the Agent decides on run-time which tables to query on. Unstructured documents can be bundled together more easily but placing all document types together can give false-positive search results ?
Mmm... Old school engineer here, spent more 30 years with db systems. And now I understand I don't want a black box RAG system but I want to implement AI stuff with PG ! For me there is a little thing you can do better: add some examples for retrieving data without and with the vectors, especially when there is a where clause. Other than that, your video is a big source of inspiration. Thanks a lot.
Thanks for the feedback!
Well, after trying the whole thing I think the caveat here is to use pgai that depends on OpenAI: this is not local, you have to pay for the tokens, your data are going away and it is a black box. So I found another way, coding some functions locally in Postgres to use ollama locally with local models: no privacy leak or data leak, no token cost. This is what I understood but I am a rookie.
pgai supports Ollama so you can use local models for greater privacy and lower cost. Check out the Github repo for more. The example used in the video is with OpenAI, but pgai also supports Ollama, Cohere, and Anthropic models too.
what is the postgres sql client being used? looks like supabase
PopSQL, it was acquired by Timescale a few months ago
It's PopSQL
Also, where is StreamingDiskANN come from? Seems only IVFFLAT and HNSW are here but diskann SQL Error [42704]: ERROR: access method "diskann" does not exist! Ha, got it: pgvectorscale !
Correct, install pgvectorscale and you can access the StreamingDiskANN index
Appreciate the video. Just FYI, you're pronouncing the word "build" like "bulled" throughout the video, but most native speakers pronounce it like "billed"
give me more videos like this please : )