Nested K/V (nested_kv_task)

This task runs K/V lookups on synthetic data. You can run it with icml_experiments/nested_kv_task/run.sh.

Document Q/A (doc_qa_task)

This task runs question answering on a set of embedded wikipedia passages.

Setup

You need a a running postgres database to run this experiment and an OpenAI account. Set your enviornment variables:

export PGVECTOR_TEST_DB_URL=postgresql+pg8000://{username}:{password}@localhost:8888/{db}
export OPENAI_API_KEY={key}

Download data

Download the wikipedia embedding at:

huggingface-cli download nlpkevinl/wikipedia_openai_embeddings --repo-type dataset

Loading embeddings

Run the script ./0_load_embeddings.sh.

This step will take a while. You can check the status of the loading by connecting to psql:

> psql -h localhost -p {password} -U {username} -d {db}
> SELECT COUNT(*) from letta_passages;

Once completed, there will be ~19 million rows in the database.

Creating an index

To avoid extremeley slow queries, you need to create an index:

CREATE INDEX ON letta_passages USING hnsw (embedding vector_l2_ops);

You can check to see if the index was created successfully with:

> SELECT indexname, indexdef FROM pg_indexes WHERE tablename = 'letta_passages';

letta_passages_embedding_idx | CREATE INDEX letta_passages_embedding_idx ON public.letta_passages USING hnsw (embedding vector_cosine_ops) WITH (m='24', ef_construction='100')

Running Document Q/A

Run the script ./1_run_docqa.sh {model_name} {n_docs} {letta/model_name}.

Evaluation

Run the script ./2_run_eval.sh.