Rottnest : Data Lake Indices

Despite our affiliations, this is not an official Anthropic-supported project! Please raise issues here on Github.

You don't need ElasticSearch or some vector database to do full text search or vector search. Parquet + Rottnest is all you need. Rottnest is like Postgres indices for Parquet. Read more on what it can do for e.g. logs here.

Installation

Local installation: pip install rottnest.

Rottnest supports many different index types. Currently you need to install separate dependencies for each index type. I also don't yet support installation like pip install rottnest[bm25] because the dependencies are still being finalized. For the time being you can:

• BM25: pip install duckdb openai if you want to rely on OpenAI embeddings for query expansion or if you want to use the FlagEmbedding library, pip install duckdb FlagEmbedding, it will be free but the quality will be lacking.
• Phrase search: Nothing more needed.
• Vector: pip install faiss (I am still working on properly supporting this, but most likely I will heavily rely on faiss to build a PQ-IVF index with SOAR).

How to use

Build indices on your Parquet files, merge them, and query them. Very simple. Let's walk through a very simple example, in demo.py. It builds a BM25 index on two Parquet files, merges the indices, and searches the merged index for records related to cell phones. The code is here:

import rottnest
rottnest.index_file_bm25("example_data/0.parquet", "body", "index0")
rottnest.index_file_bm25("example_data/1.parquet", "body", "index1")
rottnest.merge_index_bm25("merged_index", ["index0", "index1"])
result = rottnest.search_index_bm25(["merged_index"], "cell phones", K = 10)

This code will still work if the Parquet files are in fact on object storage. You can copy the data files to an S3 bucket, say s3://example_data/. Then the following code will work:

import rottnest
rottnest.index_file_bm25("s3://example_data/0.parquet", "body", "index0")
rottnest.index_file_bm25("s3://example_data/1.parquet", "body", "index1")
rottnest.merge_index_bm25("merged_index", ["index0", "index1"])
result = rottnest.search_index_bm25(["merged_index"], "cell phones", K = 10)

The indices themselves can also be on object storage.

Rottnest client will use the index to search against the Parquet files on S3 directly. Rottnest has its own Parquet reader that makes this very efficient.

If you are using S3-compatible file systems, like Ceph, MinIO, Alibaba or Volcano Cloud that might require virtual host style and different endpoint URL, you should set the following environment variables:

export AWS_ENDPOINT_URL=https://tos-s3-cn-beijing.volces.com
export AWS_VIRTUAL_HOST_STYLE=true

Rottnest not only supports BM25 indices but also other indices, like regex and vector searches. More documentation will be forthcoming.

Phrase Matches

Unlike BM25 which works on single terms, you can also build exact substring match indices which rely on the FM-index (for the moment). This is based on exact match. We are working on a Kmer-hash based method with minimizers to reduce the storage size. The code is here:

import rottnest
rottnest.index_file_substring("example_data/0.parquet", "body", "index0")
rottnest.index_file_substring("example_data/1.parquet", "body", "index1")
rottnest.merge_index_substring("merged_index", ["index0", "index1"])
result = rottnest.search_index_substring(["merged_index"], "cell phones", K = 10)

Vector Approximate Nearest Neighbor

Serverless Search Engine Architecture

Rottnest can be used to build a serverless search engine. The client will use the index to search against the Parquet files on S3 directly, or Parquet files hosted by somebody else, like Huggingface. More documentation will be forthcoming. The (simplest possible) searcher Lambda code can be found in lambda/ directory.

Development

Build Python wheel

maturin develop --features "py,opendal"

or

maturin develop --features "py,aws_sdk"