RAG & Embedding Configuration#

The RAG pipeline in Cyber Huatuo uses a local embedding model downloaded from ModelScope to vectorize documents for retrieval. All embedding settings are centralized in config/config-web.yaml under the model.embedding key, and loaded at runtime via the Config singleton.

Config File: model.embedding Keys#

Defined in config/config-web.yaml:

Note: Only model-path and model-name are read by the Python runtime. The model-version and device keys are declared in the YAML but not consumed by any code.

The knowledge base document directory is also set in the same file 1:

Knowledge-base-path: ./konwledge-base

How Configuration is Loaded#

Config is a thread-safe singleton. It selects the YAML file by reading the PY_ENVIRONMENT environment variable and opening config/config-{env}.yaml 2. For web deployments, this resolves to config-web.yaml. Nested values are accessed via get_with_nested_params(), which is memoized with @lru_cache.

To switch environments, set PY_ENVIRONMENT before launch (e.g., PY_ENVIRONMENT=web).

Where Embedding Config is Consumed#

Two model classes read from model.embedding:

1. model/RAG/retrieve_model.py — Local Knowledge Base RAG

Retrievemodel.__init__ reads both model-path and model-name, joins them to form the full local model path 3, and auto-downloads the model via snapshot_download if it is not already present on disk 4. It then initializes ModelScopeEmbeddings with the resolved path and builds a FAISS vector store from the local knowledge base files.

2. model/Internet/Internet_model.py — Internet Search RAG

InternetModel.__init__ reads only model-name (not model-path) and passes it directly as the model_id to ModelScopeEmbeddings 5. This means internet-search RAG relies on ModelScope's default resolution rather than the explicit local cache path.

RAG Pipeline Architecture#

rag/rag_chain.py
  └─> rag/retrieve/retrieve_document.py
        └─> model/RAG/retrieve_service.py
              └─> model/RAG/retrieve_model.py ← embedding config applied here

The local knowledge base retriever supports PDF, DOCX, TXT, HTML, MHTML, CSV, JSON, and Markdown files loaded from the Knowledge-base-path directory 6. Documents are chunked with RecursiveCharacterTextSplitter and indexed into FAISS.

Changing the Embedding Model#

To swap the embedding model:

1. Update model-name in config/config-web.yaml to any valid ModelScope model ID.
2. Update model-path if you want to control the download cache location 7.
3. The model is downloaded automatically on first run if the path does not exist 4.
4. Note that InternetModel ignores model-path — it uses only model-name for resolution 8.