# yaml-language-server: $schema=../references/diagram.schema.json
title: Retrieval-augmented generation with open components
subtitle: Documents are chunked and embedded into Qdrant; a FastAPI and LangChain backend retrieves context and answers with a local Ollama model or a hosted one
direction: right
# One port per edge, so the upsert and search edges into Qdrant keep their own labels.
layout: { merge: false }

legend:
  solid: Data flow
  dashed: Alternative model
  dotted: Tracing

groups:
  # Two lanes so the flows do not interleave. The embedding model and Qdrant serve both lanes, and the
  # answer models and tracing serve queries only, so all of them sit outside the lanes.
  ingestion: { type: zone, label: Ingestion }
  query: { type: zone, label: Query }

nodes:
  docs: { icon: generic:files, label: Documents, note: "PDF, HTML, Markdown", in: ingestion }
  loaders: { icon: tabler:file-import, label: Loaders, note: Parse and clean, in: ingestion }
  chunker: { icon: tabler:file-scissors, label: Chunking, note: 512-token chunks, in: ingestion }
  embedder: { icon: tabler:chart-scatter-3d, label: Embedding, note: Embed and upsert, in: ingestion }
  user: { icon: generic:user, label: User, in: query }
  react: { icon: logos:react, label: React, note: Chat UI, in: query }
  fastapi: { icon: logos:fastapi-icon, label: FastAPI, note: Chat server, in: query }
  langchain: { icon: logos:langchain-icon, label: LangChain, note: Retrieval chain, in: query }
  redis: { icon: logos:redis, label: Redis, note: Answer cache, in: query }
  # bge-m3 is a BAAI model, so it gets a neutral glyph rather than a Hugging Face mark.
  bge: { icon: tabler:brain, label: bge-m3, note: Embeds chunks and queries }
  qdrant: { icon: logos:qdrant-icon, label: Qdrant, note: Vector database }
  ollama: { icon: si:ollama, label: Ollama, note: Local LLM }
  hosted: { icon: generic:cloud, label: Hosted model API, note: OpenAI-compatible }
  # No Langfuse mark in the packs: a neutral glyph with the product name.
  langfuse: { icon: generic:monitoring, label: Langfuse, note: Traces and evals }

edges:
  - docs -> loaders -> chunker -> embedder
  - embedder -> bge: embed chunks
  - embedder -> qdrant: upsert
  - user -> react -[POST /chat]-> fastapi
  - fastapi -> langchain
  - fastapi -> redis
  - langchain -> bge: embed query
  - langchain -> qdrant: search
  - langchain -> ollama: prompt and context
  - langchain --> hosted: alternative
  - langchain ..> langfuse: traces