This guide will help you improve RAG accuracy by as much as 35% and optionally monitor your accuracy with only one line of code.

If you just want to observe your pipeline without changing behavior, check out the observe quick start guide

1. Create an Account

Go to https://www.pongo.ai/ to create an account- be sure to save the API key generated during onboarding or get one from the API keys page

2. Install the Pongo Client

pip install --upgrade pongo-python

3. Semantic Filter with Pongo

Run your pipeline and pass the top ~200 results to Pongo along with your query, then you’ll get the top-k back in order of relevance for use in your application!

You can also enable observability at this step by

import pongo

#Replace the key with your actual API key
pongo_client = pongo.PongoClient("YOUR_PONGO_KEY")


queries = ["What color are apples?", "Who made the first mobile phone?", "How many hearts do squids have?"]

#pass in the top ~100-200 results from your existing pipeline, passing more results will catch more edge cases but take slightly longer to process
#You can use objects like the first row, or raw text like the other two
lists_of_results = [
    # Apples results
    [
        {'text': 'Oranges are normally orange, unlike apples', 'metadata': {'source': 'Fruit documentation'}},
        {'text': 'Grapes can be purple or green', 'metadata': {'source': 'Fruit documentation'}},
        {'text': 'Apples can be green or red.', 'metadata': {'source': 'Fruit documentation'}},
        {'text': "If an apple is brown, it's best not to eat it.", 'metadata': {'source': 'Fruit documentation'}}
    ],
    # Mobile phone results
    [
        'Apple released the first iPhone on June 29, 2007',
        'The telephone was invented by Alexander Graham Bell in 1876.',
        'The first long-distance telephone call was made in August 1876, between Brantford and Paris, Ontario',
        'The newest iPhone models are the iPhone 15, it was released on September 22, 2023',
        'The first handheld mobile phone was the Motorola DynaTAC 8000X',
        'Martin Cooper, an engineer at Motorola, is credited with inventing the first handheld cellular mobile phone and making the first mobile phone call'
    ],
    # Squids results
    [
        'Octopuses have three hearts.',
        "A squid's systemic (main) heart has three chambers.",
        'The creature with the most hearts is the earthworm, with 10.',
        'Squids have three hearts- one systemic (main) heart and two branchial hearts'
    ]
]

for i in range(len(queries)):
    #observe=True adds automatic evaluation to queries, you'll get a regular email report and can view / download queries via the dashboard.
    filtered_result = pongo_client.filter(docs=lists_of_results[i], query=queries[i], num_results=5, observe=True, log_metadata={'source': 'Pongo Tutorial'})
    filtered_docs = filtered_result.json()

    print(f'Top answer to: {queries[i]}: {filtered_docs[0]["text"]}\n\n')

4. Check your observability results

Check out How we calculate accuracy to learn how our algorithm works
If you enabled observability at the previous step, you’ll be able to view your results on the analytics dashboard. You’ll also get your first weekly email report!

5. Set up alerts

For high-performance applications, you can use the alerts page to get notifications when queries with certain log metadata return with no relevant context. (learn more)