lazy-graph-rag · service and data flow
Hybrid seed retrieval plus an LLM-free cached concept graph built lazily from RagBase chunks
lazy-graph-rag service and data flow
The query path combines hybrid seed search with a deterministic concept graph cached by corpus fingerprint on an Atlas volume.
ATLAS COMPOSE STACK · RAG PLUGIN + WEAVIATE + NAMED CACHE VOLUME
QUERY PATH · RETRIEVAL AND CACHE CHECK RUN CONCURRENTLY
LAZY GRAPH PIPELINE · NO LIGHTRAG · NO NEO4J · NO EXTRACTION LLM
1
request
2
dispatch
3
embed
4A
4B
6
7
5A · read all RagBase chunks
5B · cache hit
5C · cache miss → build + write
8 · answer + sources + cache/index/traversal metadata
Caller
Open WebUI / API
LiteLLM
model: lazy-graph-rag
Lazy graph plugin
orchestrate · assemble · metrics
Query embedding
nomic-embed-text
Hybrid seed search
seed_k candidates
Weaviate · RagBase
hybrid seed chunks
Budgeted expansion
relevance budget · context cap
light_gen
expanded evidence → answer
Full chunk scan
read corpus state
RagBase chunks
content fingerprint input
Graph cache volume
fingerprint → cached index
Deterministic builder
concept + co-occurrence graph
MESSAGE CONTRACT, CACHE LIFECYCLE, AND TUNING
1–4 · SEED RETRIEVAL
question → embedding → hybrid seed candidates
seed_k controls graph entry points.
Uses the shared RagBase corpus.
5 · CACHE OR BUILD
all chunks → stable content fingerprint
Cache hit reuses the deterministic index.
Cache miss builds once and writes the volume.
6 · BUDGETED EXPANSION
seed nodes → relevance-budgeted neighbors
max_context_chunks bounds evidence size.
Graph is machine-oriented and query-supporting.
7–8 · GENERATE + RETURN
light_gen receives expanded source chunks.
Response exposes cache and traversal metadata.
No extraction-model call is required.