Would be curious to know what the underlying aws ec2 instance is.

Is each DB on a dedicated instance?

If not, are there per-customer iops bounds?

<https://github.com/search?q=repo%3Apostgres%2Fpostgres+%22ma...>

I agree code change is simple, but I guess the task is complex for other reasons

Comments URL: https://news.ycombinator.com/item?id=45889261

Points: 3

# Comments: 6

https://www.ubicloud.com/

Comments URL: https://news.ycombinator.com/item?id=43101430

Points: 345

# Comments: 108

I am curious how you prevent private data from getting leaked to the auto-generated public docs. I imagine this problem does not exist in open source projects, but would become an issue if not everything discussed in company's private messenger should be used as context for generating docs.

> I don’t think “get moar ram” is a good response to that particular critique.

I do not think the blog post suggested "get more ram" as a response, but happy to clarify if you could share more details!

> Indexing in Postgres is legitimately painful

Lantern is here to make the process seamless and remove most of the pain for people building LLM/AI applications. Examples:

1. We build tools to remove the guesswork of HNSW index sizing. E.g. https://lantern.dev/blog/calculator

2. We analyze typical patterns people use when building LLM apps and suggest better practices. E.g. https://lantern.dev/blog/async-embedding-tables

3. We build alerts and triggers into our cloud database that automate the discovery of many issues via heuristics.

But at least anecdotally, it made a ton of difference.

We met <200ms latency budget with Ubicloud NVMes but had to wait seconds to get an answer from the same query with GCP persistent disks or local SSDs

You can generate CLIP embeddings locally on the DB server via:

  SELECT abstract,
       introduction,
       figure1,
       clip_text(abstract) AS abstract_ai,
       clip_text(introduction) AS introduction_ai,
       clip_image(figure1) AS figure1_ai
  INTO papers_augmented
  FROM papers;

  SELECT abstract, introduction FROM papers_augmented ORDER BY clip_text(query) <=> abstract_ai LIMIT 10;

The linked library enables embedding generation for a dozen open source models and proprietary APIs (list here: <https://lantern.dev/docs/develop/generate>, and adding new ones is really easy.

ACIDRain: Concurrency-Related Attacks on Database-Backed Web Applications: http://www.bailis.org/papers/acidrain-sigmod2017.pdf

Then, during hashing you just need a constant immutable state, which effectively expands the hash space, without incurring the mutable state overhead of replacement strings strategy.

Does this only collect logs from frontend?

Or it can also collect the backend and DB latency data related to a frontend interaction?

PL/Rust is a more performant and more feature-rich alternative to PL/pgSQL, which is the traditional UDF scripting language for Postgres.

Building a vector index (or any index for that matter) inside Postgres is a more involved process and can not be done via the UDF interface, be it Rust, C or PL/pgSQL.

So, I think even if Lantern was written in Rust, it would not be a Trusted Language Extension under this definition.

You are right that there are many trade-offs between HNSW and IVFFLAT.

E.g. IVFFLAT requires there be significant amount of data in the table, before the index is created, and assumes data distribution does not change with additional inserts (since it chooses centroids during the initial creation and never updates them)

We have also generally had harder time getting high recall with IVFFLAT on vectors from embedding models such as ada-002.

There are trade-offs, some of which we will explore in later blog posts.

This post is about one thing - HNSW index creation time across two systems, at a fixed 99% recall.

But still cool to read about Armenia's tech sector here.

We run our ci/cd benchmarks on 128dim sift vectors. We have some demos using clip embeddings (512dim) and baai/bge 768 dimensional embeddings.

Generally, smaller vectors allow higher throughput and result in smaller indexes. But the effect on performance is small. Once we merge the PR implementing vector element casts to 1 and 2 byte floats, the effect of this on throughput should be even smaller.

You’re correct that IVFFLAT would be faster for your use case. However, IVFFLAT’s shortcoming is bad recall, which means less relevant results for your application. We believe that our HNSW implementation (or other indexes) can handle use cases like yours.

We currently handle a similar use-case by rerunning our index searches with exponentially increasing LIMITs and dropping the results which are not needed. Could you send us an email at support@lantern.dev? We can generate the numbers by this weekend, and get back to you with concrete results.

By the way – not sure if you saw in our blog post, if you’re using pgvector in production and switch to Lantern, we’ll help you every single step of the way. It’s very quick, and we’ll also send you some free AirPods Pro at the end of it!

Sorry for the confusion. By “produced locally” I meant “produced on your DB server” as opposed to being an API call to a third party service such as OpenAI or HuggingFace.

(But, like... doing it remotely--on the database server as part of the query plan--frankly seems kind of crazy to me, as it is going to be so slow and add a massive CPU load to what should be an I/O workload. Makes for good demos I bet, but otherwise unusable in a database context.)

It seems like you’re worried about these workflows being on the Postgres server, which may lead to performance issues.

However, if performance becomes an issue, the functions can be executed on another server. In this approach, whether or not the functions run on the Postgres server, the end user gets access to a better developing experience as all the functions they need are available within SQL.

>...this frankly shouldn't be part of the same extension We agree. These functions are already in another repository, and not part of the same extension. The repository is here: https://github.com/lanterndata/lantern_extras

btw, we did not create our own vector type and just use size-enforced real[] arrays to represent embeddings. However, you can use our index with pgvector's vector type. So, if you already have a table with pgvector's vector column type, you can start using Lantern by just creating an index on the same column.