nzachapy 0.1.3

Context filtering module that scans your data and returns only relevant portions based on a query.

nzachapy

Context Filtering for LLMs — Scan your entire dataset and return only the portion relevant to your query.

When working with large datasets and language models, stuffing everything into the context window is noisy and expensive. nzachapy solves this by chunking your data, embedding it semantically, and returning only the chunks most relevant to a given query — so your model sees signal, not noise.

---

Installation

pip install nzachapy

---

Quick Start

from nzachapy import Nzacha

# Initialize the module
nz = Nzacha()

# Add your dataset (must be a string or a list of string)
nz.add_by_words(data)

# Query for relevant context
results = nz.search("your query here")

⚠️ Important: Your dataset must be converted to a string before passing it to add_by_words().

---

Configuration

When initializing Nzacha, you can customize the following parameters. If left unset, the defaults are used.

Parameter	Default	Description
chunk_size	200	Number of words per chunk
overlap	20	Number of overlapping words between consecutive chunks
openai_api_key	None	OpenAI API key for semantic embeddings (optional)
threshold	0.6	Minimum similarity score for a chunk to be returned

Example with custom config

nz = Nzacha(
    chunk_size=150,
    overlap=30,
    openai_api_key="sk-...",
    threshold=0.75
)

---

Embedding Backends

nzachapy supports two embedding strategies:

• OpenAI — Provide your openai_api_key at initialization to use OpenAI's embedding models for higher-quality semantic search.
• Sentence Transformers (default) — If no API key is provided, the module falls back to a local sentence-transformers model. No API key required.

---

How It Works

1. Chunking — add_by_words(data) splits your string dataset into overlapping word-based chunks. chunk_size controls how many words each chunk contains; overlap controls how many words are shared between adjacent chunks to preserve context across boundaries.

2. List — add_by_words(data) if data is a list, embed each element in the list, chunk_size and overlap are not used here

3. Embedding — Each chunk/element is embedded into a vector using either OpenAI or Sentence Transformers.

4. Retrieval — When you run a query, it is embedded the same way and compared against all chunk/element embeddings. Only chunks/elements that meet or exceed the threshold similarity score are returned.

---

API Reference

Nzacha(chunk_size, overlap, openai_api_key, threshold)

Initializes the context filter with the given configuration.

nz.add_by_words(data: str | list)

if data: str Chunks the provided string dataset by word count and stores the embeddings internally else just stores the embeddings if each list element internally. Must be called before querying.

• data — Your full dataset as a string or list. Convert any other structure to a string or a list first.

---

Notes

• Always pass data as a plain string or a list. Use str(), json.dumps(), or .to_string() (for DataFrames) to convert before calling add_by_words().
• A lower threshold returns more (but less precise) chunks. A higher threshold is stricter.
• Chunk overlap helps avoid losing context that falls on chunk boundaries — tune it based on your data's structure.

---

Contributing

Contributions are welcome! If you spot an area for improvement — whether it's a new chunking strategy, a better embedding backend, performance optimizations, or additional retrieval methods — feel free to open an issue or submit a pull request.

Please make sure any changes are well-tested and clearly documented.

---

License

MIT