Graph in-Buffer Retrieval & Associative Memory

• Graph in-Buffer: Graph structure (entities + relationships) stored in RAM
• Retrieval: Query mechanism for retrieving relevant context in RAG workflows
• Associative Memory: Traverse between associated nodes via relationships, all accessed from memory

GibRAM is an in-memory knowledge graph server designed for retrieval augmented generation (RAG) workflows. It combines a lightweight graph store with vector search so that related pieces of information remain connected in memory. This makes it easier to retrieve related regulations, articles or other text when a query mentions specific subjects.

• In memory and Ephemeral: Data lives in RAM with a configurable time to live. It is meant for short lived analysis and exploration rather than persistent storage.
• Graph and Vectors Together: Stores named entities, relationships and document chunks alongside their embeddings in the same structure.
• Graph aware Retrieval: Supports traversal over entities and relations as well as semantic search, helping you pull in context that would be missed by vector similarity alone.
• Python SDK: Provides a GraphRAG style workflow for indexing documents and running queries with minimal code. Components such as chunker, extractor and embedder can be swapped out.

Install via script

curl -fsSL https://gibram.io/install.sh | sh
# Run server
gibram-server --insecure
Server runs on port 6161 by default.
# Run server
docker run -p 6161:6161 gibramio/gibram:latest
# With custom config
docker-compose up -d
pip install gibram

Basic Usage:

from gibram import GibRAMIndexer
# Initialize indexer
indexer = GibRAMIndexer(
session_id="my-project",
host="localhost",
port=6161,
llm_api_key="sk-..." # or set OPENAI_API_KEY env
)
# Index documents
stats = indexer.index_documents([
"Python is a programming language created by Guido van Rossum.",
"JavaScript was created by Brendan Eich at Netscape in 1995."]
)
print(f"Entities: {stats.entities_extracted}")
print(f"Relationships: {stats.relationships_extracted}")
# Query
results = indexer.query("Who created JavaScript?", top_k=3)
for entity in results.entities:
print(f"{entity.title}: {entity.score}")

Custom Components:

from gibram import GibRAMIndexer
from gibram.chunkers import TokenChunker
from gibram.extractors import OpenAIExtractor
from gibram.embedders import OpenAIEmbedder
indexer = GibRAMIndexer(
session_id="custom-project",
chunker=TokenChunker(chunk_size=512, chunk_overlap=50),
extractor=OpenAIExtractor(model="gpt-4o", api_key="..."),
embedder=OpenAIEmbedder(model="text-embedding-3-small", api_key="...")
)

MIT