Text Embedding Model Configuration¶

Pre-train Text Embedding Model Configuration¶

This configuration supports most pre-trained text embedding models from SentenceTransformer. Example configuration files are shown below:

all-mpnet-base-v2

gfmrag/workflow/config/text_emb_model/mpnet.yaml

_target_: gfmrag.text_emb_models.BaseTextEmbModel
text_emb_model_name: sentence-transformers/all-mpnet-base-v2
normalize: False
batch_size: 32
query_instruct: null
passage_instruct: null
model_kwargs: null

BAAI/bge-large-en

gfmrag/workflow/config/text_emb_model/bge_large_en.yaml

_target_: gfmrag.text_emb_models.BaseTextEmbModel
text_emb_model_name: BAAI/bge-large-en
normalize: True
batch_size: 32
query_instruct: "Represent this sentence for searching relevant passages: "
passage_instruct: null
model_kwargs: null

Parameter	Options	Note
`_target_`	`gfmrag.text_emb_models.BaseTextEmbModel`	The class name of Text Embedding model
`text_emb_model_name`	None	The name of the pre-train text embedding model.
`normalize`	`True`, `False`	Whether to normalize the embeddings.
`query_instruct`	None	The instruction for the query.
`passage_instruct`	None	The instruction for the passage.
`model_kwargs`	`{}`	The additional model arguments.

Qwen3 Embedding Model Configuration¶

This configuration supports Qwen3 embedding models through gfmrag.text_emb_models.Qwen3TextEmbModel. Example configuration files are shown below:

Qwen/Qwen3-Embedding-0.6B

gfmrag/workflow/config/text_emb_model/qwen3.yaml

_target_: gfmrag.text_emb_models.Qwen3TextEmbModel
text_emb_model_name: Qwen/Qwen3-Embedding-0.6B
normalize: True
batch_size: 32
query_instruct: "Instruct: Given a web search query, retrieve relevant passages that answer the query\nQuery: "
passage_instruct: null
truncate_dim: 1024

Qwen/Qwen3-Embedding-8B

gfmrag/workflow/config/text_emb_model/qwen3_8b.yaml

_target_: gfmrag.text_emb_models.Qwen3TextEmbModel
text_emb_model_name: Qwen/Qwen3-Embedding-8B
normalize: True
batch_size: 32
query_instruct: "Instruct: Given a web search query, retrieve relevant passages that answer the query\nQuery: "
passage_instruct: null
truncate_dim: 4096

Parameter	Options	Note
`_target_`	`gfmrag.text_emb_models.Qwen3TextEmbModel`	The class name of `Qwen3TextEmbModel`.
`text_emb_model_name`	`Qwen/Qwen3-Embedding-0.6B`, `Qwen/Qwen3-Embedding-8B`, or local path	The name or local path of the Qwen3 embedding model.
`normalize`	`True`, `False`	Whether to normalize the embeddings.
`batch_size`	Positive integer	Batch size used for encoding.
`query_instruct`	String, `null`	The instruction prepended to each query.
`passage_instruct`	String, `null`	The instruction prepended to each passage.
`truncate_dim`	Positive integer, `null`	Optional output embedding dimension for Matryoshka truncation, such as `1024` for `0.6B` or `4096` for `8B`.
`api_base`	URL, `null`	Base URL of an existing vLLM embedding server. If `null`, gfmrag starts a local vLLM instance.
`api_key`	String	API key used when `api_base` points to a protected server.
`vllm_timeout`	Positive integer	Timeout in seconds for vLLM embedding requests.

Nvidia Embedding Model Configuration¶

This configuration supports the Nvidia embedding models. An example of a Nvidia embedding model configuration file is shown below:

nvidia/NV-Embed-v2

gfmrag/workflow/config/text_emb_model/nv_embed_v2.yaml

_target_: gfmrag.text_emb_models.NVEmbedV2
text_emb_model_name: nvidia/NV-Embed-v2
normalize: True
batch_size: 1
query_instruct: "Instruct: Given a multi-hop question, retrieve documents that can help answer the question\nQuery: "
passage_instruct: ""
model_kwargs:
  torch_dtype: bfloat16

Parameter	Options	Note
`_target_`	`gfmrag.text_emb_models.NVEmbedV2`	The class name of `NVEmbedV2`
`text_emb_model_name`	`nvidia/NV-Embed-v2`	The name of the Nvidia embedding model.
`normalize`	`True`, `False`	Whether to normalize the embeddings.
`query_instruct`	`Instruct: Given an entity, retrieve entities that are semantically equivalent to the given entity\nQuery:`	The instruction for the query.
`passage_instruct`	None	The instruction for the passage.
`model_kwargs`	`{}`	The additional model arguments.