hybrid-rag · service and data flow
BM25 + dense relative-score fusion, optional TEI reranking, then bounded generation
hybrid-rag service and data flow
Atlas ingestion into RagBase followed by fused keyword and vector retrieval, optional TEI reranking, and answer generation.
ATLAS COMPOSE STACK · RAG TRACK
SHARED CORPUS PREPARATION
QUERY PATH
Corpus documents
selected dataset/profile
Atlas ingestion
parse · chunk · metadata
Embedding model
nomic-embed-text
Weaviate · RagBase
text index + vectors
Caller
Open WebUI / API
LiteLLM
model: hybrid-rag
Hybrid plugin
fusion · prompt · metrics
Query embedding
nomic-embed-text
Weaviate hybrid
BM25 + dense · α=0.5
TEI reranker
cross-encoder · optional
light_gen
question + top context
OpenAI response
answer · sources · metrics
1
request
2
dispatch
3
embed
4
vector
5
20 candidates
6
top 5
7
answer
rerank=false · bypass TEI and preserve fused order
MESSAGE CONTRACT AND TUNING SURFACES
0 · INDEX
RagBase stores lexical text and dense vectors.
The same index supports both retrieval signals.
No graph is constructed.
1–4 · RETRIEVAL REQUEST
question + vector → Weaviate hybrid query
alpha balances BM25 and dense relevance.
retrieve_k controls the candidate pool.
5–6 · OPTIONAL RERANK
query + candidate texts → TEI /rerank
top_n bounds the generation context.
Disabled mode bypasses TEI explicitly.
7 · GENERATE + RETURN
light_gen receives ordered evidence.
Response preserves source identifiers and scores.
Metrics expose retrieval and generation latency.