feat: Amazon Bedrock Support ⛰️

[JGalego]

This PR adds support for EMBEDDING models provided by Amazon Bedrock.

Tested with Amazon Titan Embeddings Text and Cohere Embed.

Implementation includes ad-hoc functions to add AWS SigV4 signatures.

Example:

.bail on
.mode table
.header on

.timer on

-- Load extensions
.load ./dist/debug/rembed0
.load ../sqlite-vec/dist/vec0

-- Initialize rembed clients
insert into
  temp.rembed_clients(name, options)
values 
  ('amazon.titan-embed-text-v1', 'bedrock'),
  ('amazon.titan-embed-text-v2:0', 'bedrock'),
  ('cohere.embed-multilingual-v3', 'bedrock');
  
-- Debug (extensions)
select
  *
from
  pragma_function_list
where
  builtin != 1
and (
  name like 'rembed%'
or
  name like 'vec%'
);

-- Debug (models)
with models as (
  select
	name as model_id
  from
    temp.rembed_clients
)
select
  model_id,
  length(rembed(model_id, 'where is the love?'))/4 as embedding_dimension
from
  models;

-- Debug (model + inference options)
select substr(
  hex(
    rembed(
      'cohere.embed-multilingual-v3',
	  'Where is the love?',
	  json('{"input_type": "search_query", "truncate": "END"}')
    )
  ),
  1, 100
);

-- Create articles table
create table articles(
  headline text
);

-- Populate table with random headlines
insert into articles VALUES
  ('Shohei Ohtani''s ex-interpreter pleads guilty to charges related to gambling and theft'),
  ('The jury has been selected in Hunter Biden''s gun trial'),
  ('Larry Allen, a Super Bowl champion and famed Dallas Cowboy, has died at age 52'),
  ('After saying Charlotte, a lone stingray, was pregnant, aquarium now says she''s sick'),
  ('An Epoch Times executive is facing money laundering charge');


-- Build a vector table for the embeddings
create virtual table vec_articles using vec0(
  headline_embeddings float[1536]
);

-- Embed headlines
insert into vec_articles(rowid, headline_embeddings)
  select rowid, rembed('amazon.titan-embed-text-v1', headline)
  from articles;

-- Debug
select substr(hex(headline_embeddings), 1, 100) from vec_articles;

-- Run a query
.parameter set :query 'firearm courtroom'

with matches as (
  select
    rowid,
    distance
  from vec_articles
  where headline_embeddings match rembed('amazon.titan-embed-text-v1', :query)
  order by distance
  limit 3
)
select
  headline,
  distance
from matches
left join articles on articles.rowid = matches.rowid;

Output:

+--------------------------------------------------------------+------------------+
|                           headline                           |     distance     |
+--------------------------------------------------------------+------------------+
| The jury has been selected in Hunter Biden's gun trial       | 22.7055225372314 |
+--------------------------------------------------------------+------------------+
| Shohei Ohtani's ex-interpreter pleads guilty to charges rela | 27.4765567779541 |
| ted to gambling and theft                                    |                  |
+--------------------------------------------------------------+------------------+
| Larry Allen, a Super Bowl champion and famed Dallas Cowboy,  | 29.5273017883301 |
| has died at age 52                                           |                  |
+--------------------------------------------------------------+------------------+

[JGalego]

Tested multimodal queries with Amazon Titan Multimodal Embeddings G1 (requires sqlean extensions).

🖼️ Data: images.zip

Example:

/**************
 * MULTIMODAL *
 **************/

-- Load sqlean extensions
.load ../sqlean/dist/crypto
.load ../sqlean/dist/fileio
.load ../sqlean/dist/text

-- Add Titan Embeddings Multimodal
insert into
  temp.rembed_clients(name, options)
values
  ('amazon.titan-embed-image-v1', 'bedrock');

-- Create images table
create table multimodal(
  input_text text,
  input_image text  --b64 encoded
);

-- Download images
.shell wget https://github.com/user-attachments/files/16530358/images.zip && unzip images.zip

-- Populate table with images
insert into multimodal
  select
    text_split(text_split(name, '/', -1), '.', 1),
    encode(fileio_read(name), 'base64')
  from
    fileio_ls('./images')
  where
    name like '%.jpg';

-- Build a vector table for the multimodal embeddings
create virtual table vec_multimodal using vec0(
  embeddings float[1024]
);

-- Embed text + images
insert into vec_multimodal(rowid, embeddings)
  select
    rowid,
	rembed(
      'amazon.titan-embed-image-v1',
      input_text,
      json(
        concat(
          '{"inputImage": "',
          input_image,
          '"}'
        )
      )
    )
  from
    multimodal;

-- Run a multimodal query
.parameter set :query_image './images/puppy.jpg'

with matches as (
  select
    rowid,
    distance
  from vec_multimodal
  where embeddings match
    rembed(
      'amazon.titan-embed-image-v1',
      text_split(text_split(:query_image, '/', -1), '.', 1),
      json(
        concat(
          '{"inputImage": "',
          encode(fileio_read(:query_image), 'base64'),
          '"}'
        )
      )
    )
  order by distance
  limit 3
)
select
  input_text,
  distance
from matches
left join
  multimodal
on
  multimodal.rowid = matches.rowid;

Output:

+------------+-------------------+
| input_text |     distance      |
+------------+-------------------+
| puppy      | 0.0               |
| kitty      | 0.604697287082672 |
| toy_train  | 0.807757079601288 |
+------------+-------------------+
Run Time: real 0.787 user 0.015932 sys 0.001098

[JGalego]

Added a fix for when the model ID contains 'bad' characters like colons : and forward slashes /. For a similar issue, see curl/curl#11794.

The idea is to escape the model ID once (amazon.titan-embed-text-v2:0 --> amazon.titan-embed-text-v2%3A0) to build the URL path and then escape it again to generate the canonical request and signature (amazon.titan-embed-text-v2%3A0 --> amazon.titan-embed-text-v2%253A0).

Notice that the colon can appear multiple times in the model ID e.g. cohere.embed-english-v3:0:512

[JGalego]

@asg017 would you mind reviewing this PR?