vanilla-rag · service and data flow
Dense vector retrieval with one bounded context window and one generation call
vanilla-rag service and data flow
Corpus ingestion into Weaviate and a query flow through LiteLLM, the RAG plugin, embedding, dense retrieval, and generation.
ATLAS COMPOSE STACK · RAG TRACK
CORPUS PREPARATION · ONCE PER INGESTION
QUERY PATH · ONCE PER REQUEST
Corpus documents
selected dataset/profile
Atlas ingestion
parse · chunk · metadata
Embedding model
nomic-embed-text
Weaviate · RagBase
chunk text + vectors
Caller
Open WebUI / API
LiteLLM
model: vanilla-rag
Vanilla plugin
retrieve · prompt · normalize
Query embedding
nomic-embed-text
Weaviate · RagBase
dense cosine search
light_gen
LiteLLM → configured LLM
OpenAI response
answer · sources · metrics
1
chat request
2
alias dispatch
3
question
4
nearVector · k=5
5
ranked chunks
6
answer + evidence
MESSAGE CONTRACT AND EXECUTION PHASES
0 · INDEX
chunks → embedding vectors → RagBase
No query-time graph or enrichment.
Index is reused across requests.
1–3 · DISPATCH + EMBED
POST /v1/chat/completions · model=vanilla-rag
Plugin embeds only the incoming question.
Default model properties remain alias-scoped.
4–5 · RETRIEVE
nearVector(question_vector, limit=5)
Top chunks retain source and dataset metadata.
No BM25 fusion and no reranker.
6 · GENERATE + RETURN
question + bounded retrieved context → light_gen
Plugin emits OpenAI-compatible answer payload.
Evidence and latency metrics are preserved.