apophenia 0.1.6

Extract and structure all the data from a Git repository to make them usable in RAG.

Apophenia

Apophenia give meaning to any existing Git repository.

Apophenia extract and structure all the data from a Git repository to make them usable in RAG or within AI agents.

Apophenia impose a meaningful interpretation on a nebulous stimulus (a Git repo).

Install

$ pip install apophenia

Usage

Extract data from a given repository:

$ apophenia https://github.com/4383/niet \
  --faiss_path /tmp/results.faiss \
  --metadata_path /tmp/results.json

And use generated data in a RAG (python snippet example):

import faiss
import json
import numpy as np
from sentence_transformers import SentenceTransformer
from transformers import AutoModelForCausalLM, AutoTokenizer

# Load the FAISS index and the JSON metadata previously generated
def load_index_and_metadata(faiss_path, metadata_path):
    index = faiss.read_index(faiss_path)
    with open(metadata_path, 'r', encoding='utf-8') as f:
        metadata = json.load(f)
    return index, metadata

# Embedding of the user request
def embed_query(query, model):
    return model.encode(query, convert_to_tensor=True).cpu().numpy()

# Seach in the FAISS index
def search_in_faiss(index, query_embedding, metadata, k=5):
    distances, indices = index.search(np.array([query_embedding]), k)
    results = []
    for i, idx in enumerate(indices[0]):
        result = metadata[idx]
        result['distance'] = distances[0][i]
        results.append(result)
    return results

# Build a prompt for a generative model
def build_prompt(query, retrieved_info):
    prompt = f"Answer the following question based on the retrieved information:\n\n"
    prompt += f"Question: {query}\n\n"
    prompt += "Retrieved Information:\n"
    for info in retrieved_info:
        content_type = info.get("type", "unknown")
        content_preview = info.get("content_preview", "No preview available")
        prompt += f"- {content_type.upper()}: {content_preview}\n"
    prompt += "\nYour Answer:"
    return prompt

# Generate a response with a generative model
def generate_response(prompt, model_name="EleutherAI/gpt-neo-125M", max_length=200):
    tokenizer = AutoTokenizer.from_pretrained(model_name)
    model = AutoModelForCausalLM.from_pretrained(model_name)
    input_ids = tokenizer(prompt, return_tensors="pt").input_ids
    output = model.generate(input_ids, max_length=max_length)
    return tokenizer.decode(output[0], skip_special_tokens=True)


def run_rag_system(query, faiss_path, metadata_path, embedding_model_name, generative_model_name):
    # Load data (FAISS index and metadata, and embedding)
    index, metadata = load_index_and_metadata(faiss_path, metadata_path)
    embedding_model = SentenceTransformer(embedding_model_name)

    query_embedding = embed_query(query, embedding_model)

    # Search in FAISS
    retrieved_info = search_in_faiss(index, query_embedding, metadata)

    prompt = build_prompt(query, retrieved_info)

    response = generate_response(prompt, model_name=generative_model_name)

    return response, retrieved_info

if __name__ == "__main__":
    # Configuration
    FAISS_PATH = "results.faiss"
    METADATA_PATH = "results.json"
    EMBEDDING_MODEL_NAME = "sentence-transformers/all-MiniLM-L6-v2"
    GENERATIVE_MODEL_NAME = "EleutherAI/gpt-neo-125M"

    query = "How does the authentication system work in this repository?"

    response, retrieved_info = run_rag_system(
        query=query,
        faiss_path=FAISS_PATH,
        metadata_path=METADATA_PATH,
        embedding_model_name=EMBEDDING_MODEL_NAME,
        generative_model_name=GENERATIVE_MODEL_NAME
    )

    print("Generated Response:")
    print(response)
    print("\nRetrieved Information:")
    for info in retrieved_info:
        print(info)

For more details:

$ apophenia -h

Applications

The results (FAISS vectors and JSON metadata) generated by Apophenia can be used within a RAG (Retrieval-Augmented Generation) system. Here’s a list of potential applications for using apophenia:

• Generate enriched answers by combining documentation, commit messages, and code. Example questions include: "How do I use the authenticate_user function?" or "What is the structure of this project?"
• Quickly search for specific parts of the code or documentation. Identify relevant functions or files based on queries like "Where is the authentication logic implemented?" or "Which module handles network connections?"
• Retrieve historical information to understand bugs or errors. Analyze recent changes with queries like "What are the latest modifications in this file?" or "Which commits mention this bug?"
• Automatically generate a changelog based on commit messages and diffs for a new release.
• Identify outdated dependencies or technologies and plan migrations. For instance, answer queries like "Which files are using Eventlet?" or "Which commits introduced asyncio?"
• Search for changes related to vulnerabilities or critical dependencies. Example questions include "Which files use OpenSSL?" or "Which commits fixed vulnerabilities?"
• Generate technical guides or manuals from existing code and documentation fragments. For example, create an installation guide from README files and configuration scripts.
• Understand individual contributions or file evolution by asking questions like "Who wrote this function?" or "What are John Doe's contributions?"
• Search for specific concepts within the project, such as "Where is the caching logic handled?" or "Which files mention secure connections?"
• Simplify onboarding for new developers by providing guided answers like "The main features of this project are documented in README.md." or "auth.py handles authentication logic."
• Identify which files or functions are impacted by a specific commit with questions like "Which files were modified by this commit?" or "Which tests are affected by this change?"
• Extract code examples from existing fragments in files or commits. For instance, generate a snippet to illustrate how to use a specific function or module.
• Quickly find useful information to solve a technical issue, such as "Which file is responsible for this exception?" or "Which commit introduced this error?"
• Identify the libraries used and their versions. Example questions include "Which version of Django is being used?" or "Which commits mention outdated dependencies?"
• Search for changes related to performance optimization with questions like "Which commits optimized this file?" or "Which functions were refactored for better performance?"
• Identify team members who are most active in certain areas of the project by asking "Who contributes the most to the networking module?" or "What are the primary files in this project?"
• Create customized reports on the state or evolution of a project. For example, generate a report on the 10 most significant recent commits or list the main modules and the most modified files.
• Integrate extracted data into CI/CD pipelines. For instance, identify critical files for a specific build task.
• Compare versions of files or branches using diffs and commits.
• Identify areas of the code that need documentation or refactoring by asking "Which files lack associated documentation?" or "Which commits mention suboptimal code?"

If you recognize yourself in one of these examples then Apophenia is for you:

$ pip install apophenia

Going Further with FAISS

You can use generated output FAISS with langchain or with any modern libraries like llamaindex

• https://github.com/facebookresearch/faiss
• https://pypi.org/project/faiss/

Where apophenia stands for?

Apophenia (/æpoʊˈfiːniə/) is the tendency to perceive meaningful connections between unrelated things.

Apophenia has also come to describe a human propensity to unreasonably seek definite patterns in random information, such as can occur in gambling.

https://en.wikipedia.org/wiki/Apophenia