Project 02 of 39 · AI systems · Flagship
kssrag
Retrieval-augmented generation framework published on PyPI, built around one command: point the CLI at a document and it loads, chunks, indexes, and serves a FastAPI endpoint with token-by-token SSE streaming and per-session conversation memory. Six vector stores and two retrievers are chosen by config or replaced with your own classes by import path, and any hosted or local model provider can be swapped with a single flag, with automatic fallback when one is unavailable.
01Overview
A retrieval-augmented generation framework published on PyPI, built around one idea: getting from a document to a live streaming question-answering API should take a single command.
Point the CLI at a source-of-truth document, hand it a system prompt, and it loads, chunks, indexes and serves a FastAPI endpoint with token-by-token server-sent events and per-session conversation memory.
02How it is built
- Six vector stores and two retrievers are chosen by configuration, or replaced entirely with your own classes through an import path, without forking the package.
- Any hosted or local model provider can be swapped with one flag, with automatic fallback to a backup model when one is unavailable.
- Rolling conversation memory compresses history into running summaries, so context stays bounded across a long session.
- Marker-aware buffering stops partially formed markup from reaching the client mid-stream.
- The same pipeline is available as a programmatic API and as a one-shot CLI query when a server is not needed.