streamwish 0.1.0

StreamWish: A Retrieval-Augmented Generation (RAG) library optimized for small context windows using graph traversal.

StreamWish 🌠

StreamWish is a "Streaming Graph RAG" library designed for Small Language Models (SLMs) and high-velocity data streams. It solves the "Small Context Window" problem using a Conveyor-Belt Graph and a Token Auctioneer.

Features

• Conveyor-Belt Graph: Manages a sliding window of knowledge nodes.
• Token Auctioneer: Smartly allocates limited context slots to the most valuable information (Query, Summary, Neighborhood, Path).
• Context Distillation: Automatically shrinks incoming data chunks using LLM summarization.
• Pluggable Backends: Works with OpenAI, Gemini, or local models via LLMBackend.
• Generic Ingestion: Supports .txt, .pdf, .xlsx, and Google Drive/GCP.

Why StreamWish? (VS The Big Players)

1. The "10-Slot" Philosophy Most RAG libraries (LangChain, LlamaIndex, Graphiti) optimize for finding the right answer assuming you have a 128k+ context window to fit it all in. StreamWish is built for the Edge / SLM era. It assumes you can only "see" 10 items at a time. It uses a Token Auctioneer to ruthlessly prioritize what gets into that tiny window.

2. Streaming vs. Batch "Graph RAG" (like Microsoft's implementation) usually requires processing all your documents at once to build community clusters. This is slow and expensive. StreamWish is stateful and streaming. It builds the graph node-by-node as data flows in (like a conveyor belt), making it perfect for real-time feeds, chat logs, or live sensor data.

Installation

pip install streamwish

Quick Start

from streamwish.manager import SlotManager
from streamwish.llm_interface import OpenAIAdapter

# 1. Initialize
llm = OpenAIAdapter(api_key="your-key")
manager = SlotManager(llm_client=llm)

# 2. Ingest Data
manager.ingest_stream([
    "StreamWish optimizes context for small models.",
    "It uses a graph traversal algorithm."
])

# 3. Query
response = manager.process_query("How does StreamWish handle small context?")
print(response)

License

MIT