feat: FastEmbed embedding support (#291)
Browse files### Description
Following up on https://github.com/infiniflow/ragflow/pull/275, this PR
adds support for FastEmbed model configurations.
The options are not exhaustive. You can find the full list
[here](https://qdrant.github.io/fastembed/examples/Supported_Models/).
P.S. I ran into OOM issues when building the image.
### Type of change
- [x] New Feature (non-breaking change which adds functionality)
---------
Co-authored-by: KevinHuSh <[email protected]>
- api/db/init_data.py +57 -0
- rag/llm/__init__.py +2 -1
- rag/llm/embedding_model.py +31 -0
- requirements.txt +1 -0
api/db/init_data.py
CHANGED
@@ -109,6 +109,11 @@ factory_infos = [{
|
|
109 |
"logo": "",
|
110 |
"tags": "LLM,TEXT EMBEDDING",
|
111 |
"status": "1",
|
|
|
|
|
|
|
|
|
|
|
112 |
},
|
113 |
{
|
114 |
"name": "Xinference",
|
@@ -268,6 +273,58 @@ def init_llm_factory():
|
|
268 |
"max_tokens": 128 * 1000,
|
269 |
"model_type": LLMType.CHAT.value
|
270 |
},
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
271 |
]
|
272 |
for info in factory_infos:
|
273 |
try:
|
|
|
109 |
"logo": "",
|
110 |
"tags": "LLM,TEXT EMBEDDING",
|
111 |
"status": "1",
|
112 |
+
}, {
|
113 |
+
"name": "FastEmbed",
|
114 |
+
"logo": "",
|
115 |
+
"tags": "TEXT EMBEDDING",
|
116 |
+
"status": "1",
|
117 |
},
|
118 |
{
|
119 |
"name": "Xinference",
|
|
|
273 |
"max_tokens": 128 * 1000,
|
274 |
"model_type": LLMType.CHAT.value
|
275 |
},
|
276 |
+
# ------------------------ FastEmbed -----------------------
|
277 |
+
{
|
278 |
+
"fid": factory_infos[5]["name"],
|
279 |
+
"llm_name": "BAAI/bge-small-en-v1.5",
|
280 |
+
"tags": "TEXT EMBEDDING,",
|
281 |
+
"max_tokens": 512,
|
282 |
+
"model_type": LLMType.EMBEDDING.value
|
283 |
+
}, {
|
284 |
+
"fid": factory_infos[5]["name"],
|
285 |
+
"llm_name": "BAAI/bge-small-zh-v1.5",
|
286 |
+
"tags": "TEXT EMBEDDING,",
|
287 |
+
"max_tokens": 512,
|
288 |
+
"model_type": LLMType.EMBEDDING.value
|
289 |
+
}, {
|
290 |
+
}, {
|
291 |
+
"fid": factory_infos[5]["name"],
|
292 |
+
"llm_name": "BAAI/bge-base-en-v1.5",
|
293 |
+
"tags": "TEXT EMBEDDING,",
|
294 |
+
"max_tokens": 512,
|
295 |
+
"model_type": LLMType.EMBEDDING.value
|
296 |
+
}, {
|
297 |
+
}, {
|
298 |
+
"fid": factory_infos[5]["name"],
|
299 |
+
"llm_name": "BAAI/bge-large-en-v1.5",
|
300 |
+
"tags": "TEXT EMBEDDING,",
|
301 |
+
"max_tokens": 512,
|
302 |
+
"model_type": LLMType.EMBEDDING.value
|
303 |
+
}, {
|
304 |
+
"fid": factory_infos[5]["name"],
|
305 |
+
"llm_name": "sentence-transformers/all-MiniLM-L6-v2",
|
306 |
+
"tags": "TEXT EMBEDDING,",
|
307 |
+
"max_tokens": 512,
|
308 |
+
"model_type": LLMType.EMBEDDING.value
|
309 |
+
}, {
|
310 |
+
"fid": factory_infos[5]["name"],
|
311 |
+
"llm_name": "nomic-ai/nomic-embed-text-v1.5",
|
312 |
+
"tags": "TEXT EMBEDDING,",
|
313 |
+
"max_tokens": 8192,
|
314 |
+
"model_type": LLMType.EMBEDDING.value
|
315 |
+
}, {
|
316 |
+
"fid": factory_infos[5]["name"],
|
317 |
+
"llm_name": "jinaai/jina-embeddings-v2-small-en",
|
318 |
+
"tags": "TEXT EMBEDDING,",
|
319 |
+
"max_tokens": 2147483648,
|
320 |
+
"model_type": LLMType.EMBEDDING.value
|
321 |
+
}, {
|
322 |
+
"fid": factory_infos[5]["name"],
|
323 |
+
"llm_name": "jinaai/jina-embeddings-v2-base-en",
|
324 |
+
"tags": "TEXT EMBEDDING,",
|
325 |
+
"max_tokens": 2147483648,
|
326 |
+
"model_type": LLMType.EMBEDDING.value
|
327 |
+
},
|
328 |
]
|
329 |
for info in factory_infos:
|
330 |
try:
|
rag/llm/__init__.py
CHANGED
@@ -24,7 +24,8 @@ EmbeddingModel = {
|
|
24 |
"Xinference": XinferenceEmbed,
|
25 |
"Tongyi-Qianwen": HuEmbedding, #QWenEmbed,
|
26 |
"ZHIPU-AI": ZhipuEmbed,
|
27 |
-
"Moonshot": HuEmbedding
|
|
|
28 |
}
|
29 |
|
30 |
|
|
|
24 |
"Xinference": XinferenceEmbed,
|
25 |
"Tongyi-Qianwen": HuEmbedding, #QWenEmbed,
|
26 |
"ZHIPU-AI": ZhipuEmbed,
|
27 |
+
"Moonshot": HuEmbedding,
|
28 |
+
"FastEmbed": FastEmbed
|
29 |
}
|
30 |
|
31 |
|
rag/llm/embedding_model.py
CHANGED
@@ -13,12 +13,14 @@
|
|
13 |
# See the License for the specific language governing permissions and
|
14 |
# limitations under the License.
|
15 |
#
|
|
|
16 |
from zhipuai import ZhipuAI
|
17 |
import os
|
18 |
from abc import ABC
|
19 |
from ollama import Client
|
20 |
import dashscope
|
21 |
from openai import OpenAI
|
|
|
22 |
from FlagEmbedding import FlagModel
|
23 |
import torch
|
24 |
import numpy as np
|
@@ -172,6 +174,34 @@ class OllamaEmbed(Base):
|
|
172 |
return np.array(res["embedding"]), 128
|
173 |
|
174 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
175 |
class XinferenceEmbed(Base):
|
176 |
def __init__(self, key, model_name="", base_url=""):
|
177 |
self.client = OpenAI(api_key="xxx", base_url=base_url)
|
@@ -187,3 +217,4 @@ class XinferenceEmbed(Base):
|
|
187 |
res = self.client.embeddings.create(input=[text],
|
188 |
model=self.model_name)
|
189 |
return np.array(res.data[0].embedding), res.usage.total_tokens
|
|
|
|
13 |
# See the License for the specific language governing permissions and
|
14 |
# limitations under the License.
|
15 |
#
|
16 |
+
from typing import Optional
|
17 |
from zhipuai import ZhipuAI
|
18 |
import os
|
19 |
from abc import ABC
|
20 |
from ollama import Client
|
21 |
import dashscope
|
22 |
from openai import OpenAI
|
23 |
+
from fastembed import TextEmbedding
|
24 |
from FlagEmbedding import FlagModel
|
25 |
import torch
|
26 |
import numpy as np
|
|
|
174 |
return np.array(res["embedding"]), 128
|
175 |
|
176 |
|
177 |
+
class FastEmbed(Base):
|
178 |
+
def __init__(
|
179 |
+
self,
|
180 |
+
key: Optional[str] = None,
|
181 |
+
model_name: str = "BAAI/bge-small-en-v1.5",
|
182 |
+
cache_dir: Optional[str] = None,
|
183 |
+
threads: Optional[int] = None,
|
184 |
+
**kwargs,
|
185 |
+
):
|
186 |
+
self._model = TextEmbedding(model_name, cache_dir, threads, **kwargs)
|
187 |
+
|
188 |
+
def encode(self, texts: list, batch_size=32):
|
189 |
+
# Using the internal tokenizer to encode the texts and get the total number of tokens
|
190 |
+
encodings = self._model.model.tokenizer.encode_batch(texts)
|
191 |
+
total_tokens = sum(len(e) for e in encodings)
|
192 |
+
|
193 |
+
embeddings = [e.tolist() for e in self._model.embed(texts, batch_size)]
|
194 |
+
|
195 |
+
return np.array(embeddings), total_tokens
|
196 |
+
|
197 |
+
def encode_queries(self, text: str):
|
198 |
+
# Using the internal tokenizer to encode the texts and get the total number of tokens
|
199 |
+
encoding = self._model.model.tokenizer.encode(text)
|
200 |
+
embedding = next(self._model.query_embed(text)).tolist()
|
201 |
+
|
202 |
+
return np.array(embedding), len(encoding.ids)
|
203 |
+
|
204 |
+
|
205 |
class XinferenceEmbed(Base):
|
206 |
def __init__(self, key, model_name="", base_url=""):
|
207 |
self.client = OpenAI(api_key="xxx", base_url=base_url)
|
|
|
217 |
res = self.client.embeddings.create(input=[text],
|
218 |
model=self.model_name)
|
219 |
return np.array(res.data[0].embedding), res.usage.total_tokens
|
220 |
+
|
requirements.txt
CHANGED
@@ -27,6 +27,7 @@ elasticsearch==8.12.1
|
|
27 |
elasticsearch-dsl==8.12.0
|
28 |
et-xmlfile==1.1.0
|
29 |
filelock==3.13.1
|
|
|
30 |
FlagEmbedding==1.2.5
|
31 |
Flask==3.0.2
|
32 |
Flask-Cors==4.0.0
|
|
|
27 |
elasticsearch-dsl==8.12.0
|
28 |
et-xmlfile==1.1.0
|
29 |
filelock==3.13.1
|
30 |
+
fastembed==0.2.6
|
31 |
FlagEmbedding==1.2.5
|
32 |
Flask==3.0.2
|
33 |
Flask-Cors==4.0.0
|