RagMark

RagMark helps RAG systems remember context between queries using lightweight in-memory bookmarking.

Lightweight, configurable context bookmarking for RAG pipelines.

RagMark is an in-memory layer that adds context awareness to any Retrieval-Augmented Generation system — enabling follow-up queries to continue from where they left off instead of restarting every time.

Features

• ⚡ < 1 ms decision latency
• 🧠 Context-aware retrieval (local vs global search)
• 🔌 Plug-and-play — works with any RAG pipeline
• ⚙️ Fully configurable via a single YAML file
• 🚫 Zero disruption to existing pipelines

Installation

pip install -e .

Quick Start

1. Configure (ragmark.yaml)

ragmark:
  enabled: true
  max_bookmarks: 5
  similarity_threshold: 0.75
  rag_door: pre        # pre | post | off
  debug: false

2. Use

from ragmark import RagMark, Node

rm = RagMark(config_path="ragmark.yaml")

# Get retrieval strategy
scope = rm.get_scope(query_embedding)

if scope["strategy"] == "local":
    results = vector_db.search_near(scope["node_ids"])
else:
    results = vector_db.search(query_embedding)

# Update context with results
nodes = [Node(id=r.id, embedding=r.embedding) for r in results]
rm.update(nodes, query_embedding)

Configuration Options

Key	Type	Default	Description
enabled	bool	true	Global on/off switch
max_bookmarks	int	5	Max active context nodes
similarity_threshold	float	0.75	Threshold for local search decision
rag_door	string	pre	Pipeline placement: pre/post/off
debug	bool	false	Enable debug logging

API

Method	Description
RagMark(config_path=...)	Initialise with a YAML config file
get_scope(query_embedding)	Returns strategy, confidence, and node IDs
update(nodes, query_embedding)	Update bookmarks with new retrieval results
reset()	Clear all bookmarks

Running Tests

python -m pytest tests/ -v

Author

Rishabh

Built with ❤️ as a lightweight open-source contribution to the RAG ecosystem.

License

MIT