Maintain a FAISS index for specified Datasette tables

See Semantic search answers: Q&A against documentation with GPT3 + OpenAI embeddings for background on this project.

Install this plugin in the same environment as Datasette.

datasette install datasette-faiss

This plugin creates in-memory FAISS indexes for specified tables on startup, using an IndexFlatL2 FAISS index type.

If the tables are modified after the server has started the indexes will not (yet) pick up those changes.

The tables to be indexed must have id and embedding columns. The embedding column must be a blob containing embeddings that are arrays of floating point numbers that have been encoded using the following Python function:

def encode(vector):
    return struct.pack("f" * len(vector), *vector)

You can import that function from this package like so:

from datasette_faiss import encode

You can specify which tables should have indexes created for them by adding this to metadata.json:

{
    "plugins": {
        "datasette-faiss": {
            "tables": [
                ["blog", "embeddings"]
            ]
        }
    }
}

Each table is an array listing the database name and the table name.

If you are using metadata.yml the configuration should look like this:

plugins:
  datasette-faiss:
    tables:
    - ["blog", "embeddings"]

The plugin makes four new SQL functions available within Datasette:

Returns the k nearest neighbors to the embedding found in the specified database and table. For example:

select faiss_search('blog', 'embeddings', (select embedding from embeddings where id = 3), 5)

This will return a JSON array of the five IDs of the records in the embeddings table in the blog database that are closest to the specified embedding. The returned value looks like this:

["1", "1249", "1011", "5", "10"]

You can use the SQLite json_each() function to turn that into a table-like sequence that you can join against.

Here's an example query that does that:

with related as (
  select value from json_each(
    faiss_search(
      'blog',
      'embeddings',
      (select embedding from embeddings limit 1),
      5
    )
  )
)
select * from blog_entry, related
where id = value

Takes the same arguments as above, but the return value is a JSON array of pairs, each with an ID and a score - something like this:

[
    ["1", 0.0],
    ["1249", 0.21042244136333466],
    ["1011", 0.29391372203826904],
    ["5", 0.29505783319473267],
    ["10", 0.31554925441741943]
]

Given a JSON array of floats, returns the binary embedding blob that can be used with the other functions:

select faiss_encode('[2.4, 4.1, 1.8]')
-- Returns a 12 byte blob
select hex(faiss_encode('[2.4, 4.1, 1.8]'))
-- Returns 9A991940333383406666E63F

The opposite of faiss_encode().

select faiss_decode(X'9A991940333383406666E63F')

Returns:

[2.4000000953674316, 4.099999904632568, 1.7999999523162842]

Note that floating point arithmetic results in numbers that don't quite round-trip to the exact same expected value.

This aggregate function can be used to find the k nearest neighbors to compare_embedding for each unique value of id in the table. For example:

select faiss_agg(
    id, embedding, (select embedding from embeddings where id = 3), 5
) from embeddings

Unlike the faiss_search() function, this does not depend on the per-table index that the plugin creates when it first starts running. Instead, an index is built every time the aggregation function is run.

This means that it should only be used on smaller sets of values - once you get above 10,000 or so the performance from this function is likely to become prohibitively expensive.

The function returns a JSON array of IDs representing the k rows with the closest distance scores, like this:

[1324, 344, 5562, 553, 2534]

You can use the json_each() function to turn that into a table-like sequence that you can join against.

Try an example fais_agg() query.

This is similar to the faiss_agg() aggregate function but it returns a list of pairs, each with an ID and the corresponding score - something that looks like this (if k was 2):

[[2412, 0.25], [1245, 24.25]]

Try an example fais_agg_with_scores() query.

To set up this plugin locally, first checkout the code. Then create a new virtual environment:

cd datasette-faiss
python3 -m venv venv
source venv/bin/activate

Now install the dependencies and test dependencies:

pip install -e '.[test]'

To run the tests:

pytest