feat: add Voyage AI vectorizer integration

DEVELOPMENT.md

@@ -143,6 +143,7 @@ To set up the tests:
     ENABLE_OLLAMA_TESTS=1
     ENABLE_ANTHROPIC_TESTS=1
     ENABLE_COHERE_TESTS=1
+    ENABLE_VOYAGEAI_TESTS=1
     ENABLE_VECTORIZER_TESTS=1
     ENABLE_DUMP_RESTORE_TESTS=1
     ENABLE_PRIVILEGES_TESTS=1

README.md

@@ -53,6 +53,7 @@ For other use cases, first [Install pgai](#installation) in Timescale Cloud, a p
      * [OpenAI](./docs/openai.md) - configure pgai for OpenAI, then use the model to tokenize, embed, chat complete and moderate. This page also includes advanced examples.
      * [Anthropic](./docs/anthropic.md) - configure pgai for Anthropic, then use the model to generate content.
      * [Cohere](./docs/cohere.md) - configure pgai for Cohere, then use the model to tokenize, embed, chat complete, classify, and rerank.
+     * [Voyage AI](./docs/voyageai.md) - configure pgai for Voyage AI, then use the model to embed.
   - Leverage LLMs for data processing tasks such as classification, summarization, and data enrichment ([see the OpenAI example](/docs/openai.md)).
 
 
@@ -166,6 +167,7 @@ You can use pgai to integrate AI from the following providers:
 - [Anthropic](./docs/anthropic.md)
 - [Cohere](./docs/cohere.md)
 - [Llama 3 (via Ollama)](/docs/ollama.md)
+- [Voyage AI](/docs/voyageai.md)
 
 Learn how to [moderate](/docs/moderate.md) content directly in the database using triggers and background jobs.
 

docs/vectorizer-api-reference.md

@@ -253,6 +253,7 @@ The embedding functions are:
 
 - [ai.embedding_openai](#aiembedding_openai)
 - [ai.embedding_ollama](#aiembedding_ollama)
+- [ai.embedding_voyageai](#aiembedding_voyageai)
 
 ### ai.embedding_openai
 
@@ -343,6 +344,49 @@ The function takes several parameters to customize the Ollama embedding configur
 
 A JSON configuration object that you can use in [ai.create_vectorizer](#create-vectorizers).
 
+### ai.embedding_voyageai
+
+You use the `ai.embedding_voyageai` function to use a Voyage AI model to generate embeddings.
+
+The purpose of `ai.embedding_voyageai` is to:
+- Define which Voyage AI model to use.
+- Specify the dimensionality of the embeddings.
+- Configure the model's truncation behaviour, and api key name.
+- Configure the input type.
+
+#### Example usage
+
+This function is used to create an embedding configuration object that is passed as an argument to [ai.create_vectorizer](#create-vectorizers):
+
+```sql
+SELECT ai.create_vectorizer(
+    'my_table'::regclass,
+    embedding => ai.embedding_voyageai(
+      'voyage-3-lite',
+      512,
+      truncate => false,
+      api_key_name => "TEST_API_KEY"
+    ),
+    -- other parameters...
+);
+```
+
+#### Parameters
+
+The function takes several parameters to customize the Ollama embedding configuration:
+
+| Name         | Type    | Default          | Required | Description                                                                                                                                                                                                                                                               |
+|--------------|---------|------------------|----------|---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|
+| model        | text    | -                | ✔        | Specify the name of the Voyage AI [model](https://docs.voyageai.com/docs/embeddings#model-choices) to use.                                                                                                                                                                |
+| dimensions   | int     | -                | ✔        | Define the number of dimensions for the embedding vectors. This should match the output dimensions of the chosen model.                                                                                                                                                   |
+| truncate     | boolean | true             | ✖        | Truncates the end of each input to fit within the chosen model's context length. Embedding fails (for a given chunk) if set to false and the context length is exceeded.                                                                                                  |
+| input_type   | text    | 'document'       | ✖        | Type of the input text, null, 'query', or 'document'.                                                                                                                                                                                                                     |
+| api_key_name | text    | `VOYAGE_API_KEY` | ✖        | Set the name of the environment variable that contains the Voyage AI API key. This allows for flexible API key management without hardcoding keys in the database. On Timescale Cloud, you should set this to the name of the secret that contains the Voyage AI API key. |
+
+#### Returns
+
+A JSON configuration object that you can use in [ai.create_vectorizer](#create-vectorizers).
+
 ## Formatting configuration
 
 You use the `ai.formatting_python_template` function in `pgai` to 

docs/voyageai.md

@@ -0,0 +1,182 @@
+# Use pgai with Voyage AI
+
+This page shows you how to:
+
+- [Configure pgai for Voyage AI](#configure-pgai-for-voyage-ai)
+- [Add AI functionality to your database](#usage)
+- [Follow advanced AI examples](#advanced-examples)
+
+## Configure pgai for Voyage AI
+
+Most pgai functions require an [Voyage AI API key](https://docs.voyageai.com/docs/api-key-and-installation#authentication-with-api-keys).
+
+- [Handle API keys using pgai from psql](#handle-api-keys-using-pgai-from-psql)
+- [Handle API keys using pgai from python](#handle-api-keys-using-pgai-from-python)
+
+### Handle API keys using pgai from psql
+
+The api key is an [optional parameter to pgai functions](https://www.postgresql.org/docs/current/sql-syntax-calling-funcs.html).
+You can either:
+
+* [Run AI queries by passing your API key implicitly as a session parameter](#run-ai-queries-by-passing-your-api-key-implicitly-as-a-session-parameter)
+* [Run AI queries by passing your API key explicitly as a function argument](#run-ai-queries-by-passing-your-api-key-explicitly-as-a-function-argument)
+
+#### Run AI queries by passing your API key implicitly as a session parameter
+
+To use a [session level parameter when connecting to your database with psql](https://www.postgresql.org/docs/current/config-setting.html#CONFIG-SETTING-SHELL)
+to run your AI queries:
+
+1. Set your Voyage AI key as an environment variable in your shell:
+    ```bash
+    export VOYAGE_API_KEY="this-is-my-super-secret-api-key-dont-tell"
+    ```
+1. Use the session level parameter when you connect to your database:
+
+    ```bash
+    PGOPTIONS="-c ai.voyage_api_key=$VOYAGE_API_KEY" psql -d "postgres://<username>:<password>@<host>:<port>/<database-name>"
+    ```
+
+1. Run your AI query:
+
+   `ai.voyage_api_key` is set for the duration of your psql session, you do not need to specify it for pgai functions.
+
+    ```sql
+    SELECT * FROM ai.voyageai_embed('voyage-3-lite', 'sample text to embed');
+    ```
+
+#### Run AI queries by passing your API key explicitly as a function argument
+
+1. Set your Voyage AI key as an environment variable in your shell:
+    ```bash
+    export VOYAGE_API_KEY="this-is-my-super-secret-api-key-dont-tell"
+    ```
+
+2. Connect to your database and set your api key as a [psql variable](https://www.postgresql.org/docs/current/app-psql.html#APP-PSQL-VARIABLES):
+
+      ```bash
+      psql -d "postgres://<username>:<password>@<host>:<port>/<database-name>" -v voyage_api_key=$VOYAGE_API_KEY
+      ```
+   Your API key is now available as a psql variable named `voyage_api_key` in your psql session.
+
+   You can also log into the database, then set `voyage_api_key` using the `\getenv` [metacommand](https://www.postgresql.org/docs/current/app-psql.html#APP-PSQL-META-COMMAND-GETENV):
+
+      ```sql
+      \getenv voyage_api_key VOYAGE_API_KEY
+      ```
+
+3. Pass your API key to your parameterized query:
+    ```sql
+    SELECT *
+    FROM ai.voyageai_embed('voyage-3-lite', 'sample text to embed', api_key=>$1)
+    ORDER BY created DESC
+    \bind :voyage_api_key
+    \g
+    ```
+
+   Use [\bind](https://www.postgresql.org/docs/current/app-psql.html#APP-PSQL-META-COMMAND-BIND) to pass the value of `voyage_api_key` to the parameterized query.
+
+   The `\bind` metacommand is available in psql version 16+.
+
+4. Once you have used `\getenv` to load the environment variable to a psql variable
+   you can optionally set it as a session-level parameter which can then be used explicitly.
+   ```sql
+   SELECT set_config('ai.voyage_api_key', $1, false) IS NOT NULL
+   \bind :voyage_api_key
+   \g
+   ```
+
+   ```sql
+   SELECT * FROM ai.voyageai_embed('voyage-3-lite', 'sample text to embed');
+   ```
+
+### Handle API keys using pgai from python
+
+1. In your Python environment, include the dotenv and postgres driver packages:
+
+    ```bash
+    pip install python-dotenv
+    pip install psycopg2-binary
+    ```
+
+1. Set your Voyage AI key in a .env file or as an environment variable:
+    ```bash
+    VOYAGE_API_KEY="this-is-my-super-secret-api-key-dont-tell"
+    DB_URL="your connection string"
+    ```
+
+1. Pass your API key as a parameter to your queries:
+
+    ```python
+    import os
+    from dotenv import load_dotenv
+
+    load_dotenv()
+
+    VOYAGE_API_KEY = os.environ["VOYAGE_API_KEY"]
+    DB_URL = os.environ["DB_URL"]
+
+    import psycopg2
+
+    with psycopg2.connect(DB_URL) as conn:
+        with conn.cursor() as cur:
+            # pass the API key as a parameter to the query. don't use string manipulations
+            cur.execute("SELECT * FROM ai.voyageai_embed('voyage-3-lite', 'sample text to embed', api_key=>%s)", (VOYAGE_API_KEY,))
+            records = cur.fetchall()
+    ```
+
+   Do not use string manipulation to embed the key as a literal in the SQL query.
+
+
+## Usage
+
+This section shows you how to use AI directly from your database using SQL.
+
+- [Embed](#embed): generate [embeddings](https://platform.openai.com/docs/guides/embeddings) using a
+  specified model.
+
+### Embed
+
+Generate [embeddings](https://platform.openai.com/docs/guides/embeddings) using a specified model.
+
+- Request an embedding using a specific model:
+
+    ```sql
+    SELECT ai.voyageai_embed
+    ( 'voyage-3-lite'
+    , 'the purple elephant sits on a red mushroom'
+    );
+    ```
+
+  The data returned looks like:
+
+    ```text
+                          voyageai_embed                      
+    --------------------------------------------------------
+     [0.005978798,-0.020522336,...-0.0022857306,-0.023699166]
+    (1 row)
+    ```
+
+- Pass an array of text inputs:
+
+    ```sql
+    SELECT ai.voyageai_embed
+    ( 'voyage-3-lite'
+    , array['Timescale is Postgres made Powerful', 'the purple elephant sits on a red mushroom']
+    );
+    ```
+
+- Specify the input type
+
+  The Voyage AI API allows setting the `input_type` to `"document"`, or
+  `"query"`, (or unset). Correctly setting this value should enhance retrieval
+  quality:
+
+    ```sql
+    SELECT ai.voyageai_embed
+    ( 'voyage-3-lite'
+    , 'A query'
+    , input_type => 'query'
+    );
+    ```
+
+

projects/extension/ai/voyageai.py

@@ -0,0 +1,22 @@
+import voyageai
+from typing import Optional, Generator, Union
+
+DEFAULT_KEY_NAME = "VOYAGE_API_KEY"
+
+
+def embed(
+    model: str,
+    input: Union[list[str]],
+    api_key: str,
+    input_type: Optional[str] = None,
+    truncation: Optional[bool] = None,
+) -> Generator[tuple[int, list[float]], None, None]:
+    client = voyageai.Client(api_key=api_key)
+    args = {}
+    if truncation is not None:
+        args["truncation"] = truncation
+    response = client.embed(input, model=model, input_type=input_type, **args)
+    if not hasattr(response, "embeddings"):
+        return None
+    for idx, obj in enumerate(response.embeddings):
+        yield idx, obj

projects/extension/requirements.txt

@@ -3,4 +3,5 @@ tiktoken==0.7.0
 ollama==0.2.1
 anthropic==0.29.0
 cohere==5.5.8
-backoff==2.2.1
+backoff==2.2.1
+voyageai==0.3.1

projects/extension/setup.cfg

@@ -13,3 +13,4 @@ install_requires =
     anthropic==0.29.0
     cohere==5.5.8
     backoff==2.2.1
+    voyageai==0.3.1

projects/extension/sql/idempotent/008-embedding.sql

@@ -47,6 +47,37 @@ $func$ language sql immutable security invoker
 set search_path to pg_catalog, pg_temp
 ;
 
+-------------------------------------------------------------------------------
+-- embedding_voyageai
+create or replace function ai.embedding_voyageai
+( model text
+, dimensions int
+, truncate boolean default true
+, input_type text default 'document'
+, api_key_name text default 'VOYAGE_API_KEY'
+) returns jsonb
+as $func$
+begin
+    if input_type is not null and input_type not in ('query', 'document') then
+        -- Note: purposefully not using an enum here because types make life complicated
+        raise exception 'invalid input_type for voyage ai "%"', input_type;
+    end if;
+
+    return json_object
+    ( 'implementation': 'voyageai'
+    , 'config_type': 'embedding'
+    , 'model': model
+    , 'dimensions': dimensions
+    , 'truncate': truncate
+    , 'input_type': input_type
+    , 'api_key_name': api_key_name
+    absent on null
+    );
+end
+$func$ language plpgsql immutable security invoker
+set search_path to pg_catalog, pg_temp
+;
+
 -------------------------------------------------------------------------------
 -- _validate_embedding
 create or replace function ai._validate_embedding(config jsonb) returns void
@@ -69,6 +100,8 @@ begin
             -- ok
         when 'ollama' then
             -- ok
+        when 'voyageai' then
+            -- ok
         else
             if _implementation is null then
                 raise exception 'embedding implementation not specified';

projects/extension/sql/idempotent/015-voyageai.sql

@@ -0,0 +1,48 @@
+-------------------------------------------------------------------------------
+-- voyageai_embed
+-- generate an embedding from a text value
+-- https://docs.voyageai.com/reference/embeddings-api
+create or replace function ai.voyageai_embed
+( model text
+, input_text text
+, input_type text default null
+, api_key text default null
+, api_key_name text default null
+) returns @extschema:vector@.vector
+as $python$
+    #ADD-PYTHON-LIB-DIR
+    import ai.voyageai
+    import ai.secrets
+    api_key_resolved = ai.secrets.get_secret(plpy, api_key, api_key_name, ai.voyageai.DEFAULT_KEY_NAME, SD)
+    for tup in ai.voyageai.embed(model, [input_text], api_key=api_key_resolved):
+        return tup[1]
+$python$
+language plpython3u immutable parallel safe security invoker
+set search_path to pg_catalog, pg_temp
+;
+
+-------------------------------------------------------------------------------
+-- voyageai_embed
+-- generate embeddings from an array of text values
+-- https://docs.voyageai.com/reference/embeddings-api
+create or replace function ai.voyageai_embed
+( model text
+, input_texts text[]
+, api_key text default null
+, api_key_name text default null
+, input_type text default null
+) returns table
+( "index" int
+, embedding @extschema:vector@.vector
+)
+as $python$
+    #ADD-PYTHON-LIB-DIR
+    import ai.voyageai
+    import ai.secrets
+    api_key_resolved = ai.secrets.get_secret(plpy, api_key, api_key_name, ai.voyageai.DEFAULT_KEY_NAME, SD)
+    for tup in ai.voyageai.embed(model, input_texts, api_key=api_key_resolved):
+        yield tup
+$python$
+language plpython3u immutable parallel safe security invoker
+set search_path to pg_catalog, pg_temp
+;