Skip to main content

RAGLight 4

RAGLight is a small Python framework for applications that need control over how documents are processed, retrieved, reranked, and supplied to a language model. Version 4 is a breaking API redesign: components are injected directly and optional integrations live behind explicit extras.

Implemented capabilities

Classic RAG

Deterministic retrieval, optional reranking, generation, and streaming.

Agentic RAG

A separate bounded runtime with custom agents, custom tools, and MCP tools.

Structured results

Sources, stage events, timings, metadata, and trace IDs accompany answers.

Conversation isolation

Explicit history or caller-owned sessions; pipelines hold no shared chat state.

Conversational retrieval

Opt-in query reformulation resolves follow-up references before retrieval.

Native observability

Structured results plus optional native Langfuse traces for classic and Agentic RAG.

Directory ingestion

Recursive discovery, extension-based processing, safe skips, and per-file failures.

Flexible retrieval

Chroma or Qdrant with semantic, BM25, or hybrid retrieval.

Provider choice

Local and hosted LLMs and embedding models through optional integrations.

Structural extension points

Custom components satisfy protocols without framework inheritance.

Design principles

  • Small core: importing raglight does not import provider SDKs.
  • Explicit ownership: applications select components and own credentials.
  • Separate runtimes: classic and Agentic RAG are different APIs.
  • Observable execution: successful generation exposes structured execution data.
  • Safe state: conversation history belongs to a session or caller.
  • Composable integrations: providers can be replaced without changing pipeline code.

Framework scope

RAGLight 4 currently ships the Python framework. A command-line application, HTTP server, and dedicated user interface are delivery layers planned separately; they are not part of the current package. The framework remains usable without any delivery application.
Version 4 does not provide compatibility aliases for the previous public API. Follow the current guides and imports in this documentation.