llama-index-graph-stores-synapcores 0.1.0

LlamaIndex property graph store integration for SynapCores — for GraphRAG.

llama-index-graph-stores-synapcores

LlamaIndex property graph store integration for SynapCores. Use it as the backend for PropertyGraphIndex and GraphRAG workflows.

pip install llama-index llama-index-graph-stores-synapcores

Quickstart

from llama_index.core import PropertyGraphIndex
from llama_index.core.readers import SimpleDirectoryReader
from llama_index.graph_stores.synapcores import SynapCoresPropertyGraphStore

# 1. docker run -p 8080:8080 -e AIDB_ACCEPT_LICENSE=1 synapcores/community:latest
graph_store = SynapCoresPropertyGraphStore(
    uri="http://localhost:8080",
    graph_name="my_kb",
    embedding_dim=1536,
)

docs = SimpleDirectoryReader("./data").load_data()
index = PropertyGraphIndex.from_documents(
    docs,
    property_graph_store=graph_store,
)

retriever = index.as_retriever()
for node in retriever.retrieve("Who founded Acme and when?"):
    print(node.text)

Configuration

Argument	Default	Env fallback
uri	http://localhost:8080	SYNAPCORES_URI
database	default	SYNAPCORES_DATABASE
auth_token	None	SYNAPCORES_AUTH_TOKEN
graph_name	llama_graph	—
embedding_dim	1536	—
overwrite	False	—

graph_name tags every node + relation so multiple logical graphs can coexist in one engine without colliding.

Storage model

LlamaIndex type	Cypher label	Notes
EntityNode (name + type, e.g. "PERSON Alice")	:Entity {id, name, label, properties_json, graph_name, embedding?}	Embedding stored when provided
ChunkNode (source text)	:Chunk {id, text, label, properties_json, graph_name, embedding?}
Relation (typed edge with properties)	[:RELATION {label, properties_json, graph_name}]	label is the relation type — preserved as a property so filtering by type doesn't need to parse the edge label

Free-form properties are JSON-serialized into a single properties_json field on each node/relation. Round-tripping is reliable across every engine version (>=v1.7.0.2-ce) the package supports.

What works in v0.1.0

• upsert_nodes, upsert_relations — bulk MERGE
• get, get_triplets — filtered fetch by id / name / property
• get_rel_map(graph_nodes, depth=2, limit=30, ignore_rels=…) — depth-bounded BFS expansion, the load-bearing GraphRAG retrieval primitive
• structured_query(cypher, param_map) — pass-through Cypher with named-bind translation ($entity → $1)
• vector_query(VectorStoreQuery) — cosine similarity over Chunk-node embeddings (client-side; see Performance below)
• delete(entity_names, relation_names, properties, ids) — selective delete; refuses to wipe the entire graph without at least one selector
• aupsert_nodes, aupsert_relations, aget_*, astructured_query, avector_query — async via asyncio.to_thread

Performance

vector_query in v0.1.0 pulls every Chunk's embedding and scores client-side — O(N) in chunk count, fine for graphs up to ~thousands of chunks. The v0.2.0 path stores chunk embeddings in a parallel SQL table with HNSW for sub-linear lookup. Native get_rel_map via variable-length path matches (MATCH p = (a)-[*1..N]-(b)) is also queued for v0.2.0 — current path is single-hop Cypher in a Python BFS.

Engine requirements

• synapcores/community:v1.7.0.2-ce or newer (needs Cypher on /v1/query/execute — fixed in #223)

License

MIT