AI inference functions v7

AIDB provides a set of SQL functions for performing AI inference directly in queries. These functions let you encode data into vector representations, generate text with language models, and control inference behavior — without setting up a pipeline.

aidb.encode_text()

Encodes a single text string into a vector using a registered embedding model.

SELECT aidb.encode_text('my_embedder','What is pgvector?');

                         encode_text
--------------------------------------------------------------
 [0.021,-0.043,0.117, ...]
(1 row)

For the full parameter reference, see Models reference.

aidb.encode_text_batch()

Encodes multiple text strings into vectors in a single batched request. More efficient than calling encode_text() in a loop when processing many inputs.

SELECT * FROM aidb.encode_text_batch(
    'my_model_name',
    ARRAY['What is pgvector?', 'How does RAG work?', 'Explain embeddings.']
);

  encode_text_batch
-------------------------
 [0.021,-0.043,0.117,...]
 [0.008, 0.091,-0.034,...]
 [0.055,-0.012, 0.076,...]
(3 rows)

For the full parameter reference, see Models reference.

aidb.encode_image()

Encodes a single image into a vector using a registered multimodal embedding model (for example, a CLIP model).

SELECT aidb.encode_image(
    'my_model_name',
    pg_read_binary_file('/path/to/image.png')::BYTEA
);

                         encode_image
--------------------------------------------------------------
 [-0.012, 0.084, 0.031, ...]
(1 row)

For the full parameter reference, see Models reference.

aidb.decode_text()

Generates text from a language model (LLM) given a prompt. This is the core function for LLM inference in AIDB.

SELECT aidb.decode_text(
    model_name => 'my_model_name',
    input   => 'Explain what a knowledge base is in two sentences.',
    inference_config => NULL
);

                                        decode_text
--------------------------------------------------------------------------------------------
 A knowledge base is a structured repository of information used to answer queries or ...
(1 row)

To control generation behavior, pass an inference_config object built by aidb.inference_config(). Cast its JSONB result to json so the call matches the aidb.decode_text(TEXT, TEXT, json) signature:

SELECT aidb.decode_text(
    model_name       => 'my_model_name',
    input            => 'Summarize the following text: ...',
    inference_config => aidb.inference_config(
        system_prompt => 'You are a concise technical writer.',
        temperature   => 0.2,
        max_tokens    => 256
    )::json
);

For the full parameter reference, see Models reference.

aidb.decode_text_batch()

Generates text responses for multiple prompts in a single batched request.

SELECT aidb.decode_text_batch(
    model_name => 'my_model_name',
    input  => ARRAY[
        'What is a vector database?',
        'What is retrieval-augmented generation?'
    ]
);

                          decode_text_batch
----------------------------------------------------------
 A vector database stores high-dimensional vectors ...
 Retrieval-augmented generation (RAG) combines a ...
(2 rows)

For the full parameter reference, see Models reference.

aidb.inference_config()

Builds an inference configuration object to control language model generation behavior. Pass the result as the inference_config argument of aidb.decode_text() or aidb.decode_text_batch(), or as part of aidb.summarize_text_config() inside a pipeline step.

SELECT aidb.decode_text(
    model_name       => 'my_llm',
    input            => 'Write a haiku about PostgreSQL.',
    inference_config => aidb.inference_config(
        temperature => 0.9,
        max_tokens  => 64,
        seed        => 42
    )::json
);

For the full parameter reference, see Models reference.