contextual-rag · service and data flow
Ingestion-time chunk enrichment followed by hybrid retrieval, optional reranking, and generation
contextual-rag service and data flow
Each chunk is enriched with a model-generated document context before embedding, then queried through the contextual collection.
ATLAS COMPOSE STACK · RAG TRACK
CONTEXTUAL INGESTION · ENRICH ONCE PER CHUNK
QUERY PATH · REUSE ENRICHED INDEX
Corpus documents
source files
Atlas ingestion
parse · chunk · metadata
Weaviate · RagBase
completed plain chunks
Showcase post-step
read chunks + source docs
contextual_blurb
document window + chunk
Enrich + embed
prefix + chunk → nomic
RagContextual
enriched text + vectors
Caller
Open WebUI / API
LiteLLM
model: contextual-rag
Contextual plugin
retrieve · prompt · metrics
Query embedding
nomic-embed-text
RagContextual
hybrid search
TEI reranker
optional cross-encoder
light_gen
enriched evidence + question
OpenAI response
answer · sources · metrics
1
request
2
dispatch
3
embed
4
vector
5
candidates
6
top context
7
answer
MESSAGE CONTRACT AND EXECUTION PHASES
0A · ENRICH
RagBase chunk + source window → contextual_blurb
Showcase post-step owns the derived collection.
Cost is paid after Atlas ingestion, not per query.
0B · INDEX
context prefix + chunk → embedding
RagContextual is distinct from RagBase.
Re-ingestion is required after prompt/model changes.
1–6 · RETRIEVE + RERANK
question vector → contextual hybrid candidates
Optional TEI reranking bounds top_n context.
Retrieved text includes the generated prefix.
7 · GENERATE + RETURN
light_gen consumes enriched evidence.
Response preserves source and dataset metadata.
No graph is created or traversed.