graph-rag · LightRAG service and data flow
Persistent entity/relationship extraction with profile-driven graph, vector, and chunk retrieval
graph-rag LightRAG service and data flow
Documents are extracted into persistent LightRAG graph and vector stores, then queried through keyword and answer model roles.
ATLAS COMPOSE STACK · LIGHTRAG SERVICE + RAG PLUGIN
INDEXING · PERSISTENT KNOWLEDGE CONSTRUCTION
QUERY · PROFILE-SELECTED MODE AND BUDGETS
LIGHTRAG PERSISTENT STORES
optional TEI adapter bypass
8 · answer-only OpenAI response returns through plugin and LiteLLM
Full documents
Atlas ingestion job
LightRAG /documents
chunk + extraction orchestration
EXTRACT model role
entities · relations · summaries
Neo4j
entity/relation graph
Vector storage
entities · relations · chunks
KV / doc status
text units · pipeline state
Graph/vector lookup
profile-bounded evidence
Caller
Open WebUI / API
LiteLLM
graph-rag profile alias
Graph plugin
select profile · normalize
LightRAG /query
local · global · hybrid · mix
KEYWORD role
query decomposition
TEI adapter
optional rerank
QUERY model role
evidence → answer
1
request
2
profile alias
3
profile params
4
/query
5
keywords
6
contexts
7
ranked
MESSAGE CONTRACT, STATE, AND TUNING
0 · EXTRACT + PERSIST
full docs → chunks → entities/relationships
Graph is created during ingestion and reused.
Rebuild when extraction inputs/config change.
1–4 · PROFILE DISPATCH
LiteLLM alias identifies a query profile.
Mode, top-k, chunk-k, and token budgets vary.
The profiles share one ingested workspace.
5–7 · RETRIEVE + RERANK
LightRAG selects graph/vector evidence internally.
No showcase-controlled fixed k-hop traversal.
The query profile selects optional TEI reranking.
8 · GENERATE + RETURN
QUERY role synthesizes the final answer.
Exact retrieved contexts are not exposed.
Plugin returns answer and service metrics.