agentic-rag · service and data flow
Request-local ReAct controller chooses vector or graph tools, observes results, and emits a trace
agentic-rag service and data flow
An agent controller selects vector and graph retrieval tools during each request and keeps observations only in the current message history.
ATLAS COMPOSE STACK · RAG PLUGIN + RETRIEVAL SERVICES
CONTROL PATH · ONE REQUEST-LOCAL MESSAGE HISTORY
TOOL EXECUTION · SELECTED AT QUERY TIME
7 · finish → final answer + trace
6 · tool observations append to current request history; repeat until finish or MAX_STEPS=4
Caller
Open WebUI / API
LiteLLM
model: agentic-rag
Agentic plugin
schemas · loop · trace
ReAct controller LLM
choose tool or finish
OpenAI response
answer · sources · trace
1
request
2
dispatch
3
tools + prompt
search_vectors
query · vector_top_k
Embedding model
nomic-embed-text
Weaviate
hybrid vector search
query_graph
question · graph_mode
LightRAG
graph/vector retrieval
Tool observations
results or failure text
MESSAGE CONTRACT, STATE, AND FAILURE SEMANTICS
1–3 · INITIAL DECISION
question + tool schemas → controller model
Tool selection happens at query time.
No cross-request memory is persisted.
4 · VECTOR TOOL
question → embed → Weaviate hybrid search
Returns source-bearing chunk observations.
Uses the existing RagBase index.
5 · GRAPH TOOL
question → LightRAG query service
Uses the persistent LightRAG workspace.
Tool failures become observations, not crashes.
6–7 · LOOP + RETURN
Observations feed the next controller turn.
Loop stops at finish or MAX_STEPS.
Response includes the request execution trace.