Ash Framework•2mo ago

Following the guide in AshAi

I get this error when I am trying to use the cosine distance between the vector and the embedded question, I have the same code as the guide

function: :vector_cosine_distance, arity: nil, resource: Taleed.Poetry.Verse, could_be_calculation?: false, splode: Ash.Error, bread_crumbs: [\"Error returned from: Taleed.Poetry.Verse.vector_search\"], vars: [], path: [:filter], stacktrace: #Splode.Stacktrace<>, class: :invalid}" while retrieving Exception.message/1 for %Ash.Error.Query.NoSuchFunction{function: :vector_cosine_distance, arity: nil, resource: Taleed.Poetry.Verse, could_be_calculation?: false, splode: Ash.Error

function: :vector_cosine_distance, arity: nil, resource: Taleed.Poetry.Verse, could_be_calculation?: false, splode: Ash.Error, bread_crumbs: [\"Error returned from: Taleed.Poetry.Verse.vector_search\"], vars: [], path: [:filter], stacktrace: #Splode.Stacktrace<>, class: :invalid}" while retrieving Exception.message/1 for %Ash.Error.Query.NoSuchFunction{function: :vector_cosine_distance, arity: nil, resource: Taleed.Poetry.Verse, could_be_calculation?: false, splode: Ash.Error

36 Replies

ZachDaniel•2mo ago

Add "vector" extension and follow these steps: https://hexdocs.pm/ash_postgres/AshPostgres.Extensions.Vector.html If you get a chance, a PR to enhance that guide would be great.

Abu kumathraOP•2mo ago

I did that

ZachDaniel•2mo ago

🤔

Abu kumathraOP•2mo ago

I added the pgvector extension

ZachDaniel•2mo ago

can I see how you're calling it?

Abu kumathraOP•2mo ago

    read :vector_search do
      argument :query, :string, allow_nil?: false

      prepare before_action(fn query, context ->
                case Taleed.OpenAIEmbeddingModel.generate([query.arguments.query], []) do
                  {:ok, [search_vector]} ->
                    Ash.Query.filter(
                      query,
                      vector_cosine_distance(full_text_vector, ^search_vector) < 0.5
                    )
                    |> Ash.Query.sort(
                      {calc(vector_cosine_distance(full_text_vector, ^search_vector),
                         type: :float
                       ), :asc}
                    )
                    |> Ash.Query.limit(10)

                  {:error, error} ->
                    {:error, error}
                end
              end)
    end

    read :vector_search do
      argument :query, :string, allow_nil?: false

      prepare before_action(fn query, context ->
                case Taleed.OpenAIEmbeddingModel.generate([query.arguments.query], []) do
                  {:ok, [search_vector]} ->
                    Ash.Query.filter(
                      query,
                      vector_cosine_distance(full_text_vector, ^search_vector) < 0.5
                    )
                    |> Ash.Query.sort(
                      {calc(vector_cosine_distance(full_text_vector, ^search_vector),
                         type: :float
                       ), :asc}
                    )
                    |> Ash.Query.limit(10)

                  {:error, error} ->
                    {:error, error}
                end
              end)
    end

ZachDaniel•2mo ago

🤔 does that resource have AshPostgres.DataLayer as its data layer?

Abu kumathraOP•2mo ago

yes and I have the AshAi, AshOban extensions

ZachDaniel•2mo ago

That is very very strange...

Abu kumathraOP•2mo ago

yes it is the thing is that I copy pasted it from the guide

ZachDaniel•2mo ago

🤔 right. What if you comment out the sort/filter

Abu kumathraOP•2mo ago

I created the index and used the vectorize block

ZachDaniel•2mo ago

is it just one of them breaking it?

Abu kumathraOP•2mo ago

tried it with one of them off and it still failed I commented the filter/sort same thing

ZachDaniel•2mo ago

wait with both commented out you still get that error? or just with one or the other

Abu kumathraOP•2mo ago

no only when there is one of them from the message it says it can't find the vector_cosine_distance

ZachDaniel•2mo ago

Is this something you can push up somewhere we can check out?

Abu kumathraOP•2mo ago

I don't think I will be able to but I will try to check if I can get it to work using PSQL

ZachDaniel•2mo ago

oh, wiat lemme see your repo

Abu kumathraOP•2mo ago

defmodule Taleed.Repo do
  use AshPostgres.Repo,
    otp_app: :taleed

  @impl true
  def installed_extensions do
    # Add extensions here, and the migration generator will install them.
    ["ash-functions", "citext", "pgvector"]
  end

  # Don't open unnecessary transactions
  # will default to `false` in 4.0
  @impl true
  def prefer_transaction? do
    false
  end

  @impl true
  def min_pg_version do
    %Version{major: 17, minor: 5, patch: 0}
  end
end

defmodule Taleed.Repo do
  use AshPostgres.Repo,
    otp_app: :taleed

  @impl true
  def installed_extensions do
    # Add extensions here, and the migration generator will install them.
    ["ash-functions", "citext", "pgvector"]
  end

  # Don't open unnecessary transactions
  # will default to `false` in 4.0
  @impl true
  def prefer_transaction? do
    false
  end

  @impl true
  def min_pg_version do
    %Version{major: 17, minor: 5, patch: 0}
  end
end

I just added pgvector rn and recompiling now it works still no vectors match but no errors I get a not found so that's fixed thanks a lot. it's been a roller coaster (first time using it) using the library but the support has been great @Zach Daniel now there is a weird issue if I add pgvector to the extension it will force me to migrate and this won't work the user doesn't have super user privileges, and if I remove it I can't search do I need to install the elixir pgvector package

ZachDaniel•2mo ago

It's just called "vector" Not "pgvector" Does our guide say "pgvector"?

Abu kumathraOP•2mo ago

the guide doesn't mention anything ok so I changed it to vector and created the postgres_types.ex file as per the steps mentioned here, the issue is that there is no mention that this is required, also I installed AshAi chat using igniter

ZachDaniel•2mo ago

Gotcha, can you please open an issue or PR describing what was missing? This will fix it for the next folks 🙇‍♂️

Abu kumathraOP•2mo ago

I will do that, I will be traveling this weekend, I will do it in monday one last question does ashai create a column to mark whether the vector contains stale embedding

Abu kumathraOP•2mo ago

because it's been 2 hours and 300k embedding tasks, and the vectors are still being embeded

Abu kumathraOP•2mo ago

the issue is that the vectors were embedded, I am afraid to look at my openai bill

ZachDaniel•2mo ago

😅it should be scheduling embeddings whenever the values change You should of course trial these things before running them on real data...

Abu kumathraOP•2mo ago

I know it shouldn't, anyway thank you great library. I will write a PR+Issue on monday for the guide. Is there something in the roadmap about including real vector dbs currently postgres is complaining about the vector size if I use open ai's text-embedding-3-large which is 3072 Also Can a resource have multiple vectorize statements? I tried but failed miserably. I will probably keep a list of items that I missed from the guide so new comers won't fall into them

ZachDaniel•2mo ago

only one vectorize statement but you can vectorize multiple attributes vectorize foo: :bar, baz: :buz IIRC for doing it without pgvector everything gets more complicated but also: you can definitely do it with a vector size that big I would always prefer to use pgvector if possible before reaching for a "real" vector db

Abu kumathraOP•2mo ago

can you expand on this

ZachDaniel•2mo ago

Sorry, its like this:

vectorize do
  attributes name: :vectorized_name, biography: :vectorized_biography

  ...
end

vectorize do
  attributes name: :vectorized_name, biography: :vectorized_biography

  ...
end

so you can vectorize multiple attrs Maybe I didn't understand what you wanted though

Abu kumathraOP•2mo ago

how do I define the text that will be vectorized for each one

ZachDaniel•2mo ago

🤔 what do you mean? That takes the name attribute, and vectorizes it as vectorized_name so you'd just set the name attribute

Abu kumathraOP•2mo ago

oh that will do it actually I thought that I need to define full_text function for each attribute so I can still use the full_text one if I want to concat multiple values but I can only have one full_text function

ZachDaniel•2mo ago

Ah, yeah currently

Abu kumathraOP•2mo ago

ok thank you

Gaming

Programming

Following the guide in AshAi

Did you find this page helpful?