Kevin Hu
commited on
Commit
·
1dc3f10
1
Parent(s):
cd19d72
Refactor for total_tokens. (#4652)
Browse files### What problem does this PR solve?
#4567
### Type of change
- [x] Bug Fix (non-breaking change which fixes an issue)
- rag/llm/chat_model.py +38 -34
- rag/llm/embedding_model.py +28 -16
- rag/llm/rerank_model.py +13 -2
rag/llm/chat_model.py
CHANGED
|
@@ -53,7 +53,7 @@ class Base(ABC):
|
|
| 53 |
ans += LENGTH_NOTIFICATION_CN
|
| 54 |
else:
|
| 55 |
ans += LENGTH_NOTIFICATION_EN
|
| 56 |
-
return ans, response
|
| 57 |
except openai.APIError as e:
|
| 58 |
return "**ERROR**: " + str(e), 0
|
| 59 |
|
|
@@ -75,15 +75,11 @@ class Base(ABC):
|
|
| 75 |
resp.choices[0].delta.content = ""
|
| 76 |
ans += resp.choices[0].delta.content
|
| 77 |
|
| 78 |
-
|
| 79 |
-
|
| 80 |
-
|
| 81 |
-
+ num_tokens_from_string(resp.choices[0].delta.content)
|
| 82 |
-
)
|
| 83 |
-
elif isinstance(resp.usage, dict):
|
| 84 |
-
total_tokens = resp.usage.get("total_tokens", total_tokens)
|
| 85 |
else:
|
| 86 |
-
total_tokens =
|
| 87 |
|
| 88 |
if resp.choices[0].finish_reason == "length":
|
| 89 |
if is_chinese(ans):
|
|
@@ -97,6 +93,17 @@ class Base(ABC):
|
|
| 97 |
|
| 98 |
yield total_tokens
|
| 99 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 100 |
|
| 101 |
class GptTurbo(Base):
|
| 102 |
def __init__(self, key, model_name="gpt-3.5-turbo", base_url="https://api.openai.com/v1"):
|
|
@@ -182,7 +189,7 @@ class BaiChuanChat(Base):
|
|
| 182 |
ans += LENGTH_NOTIFICATION_CN
|
| 183 |
else:
|
| 184 |
ans += LENGTH_NOTIFICATION_EN
|
| 185 |
-
return ans, response
|
| 186 |
except openai.APIError as e:
|
| 187 |
return "**ERROR**: " + str(e), 0
|
| 188 |
|
|
@@ -212,14 +219,11 @@ class BaiChuanChat(Base):
|
|
| 212 |
if not resp.choices[0].delta.content:
|
| 213 |
resp.choices[0].delta.content = ""
|
| 214 |
ans += resp.choices[0].delta.content
|
| 215 |
-
|
| 216 |
-
|
| 217 |
-
|
| 218 |
-
|
| 219 |
-
|
| 220 |
-
if not hasattr(resp, "usage")
|
| 221 |
-
else resp.usage["total_tokens"]
|
| 222 |
-
)
|
| 223 |
if resp.choices[0].finish_reason == "length":
|
| 224 |
if is_chinese([ans]):
|
| 225 |
ans += LENGTH_NOTIFICATION_CN
|
|
@@ -256,7 +260,7 @@ class QWenChat(Base):
|
|
| 256 |
tk_count = 0
|
| 257 |
if response.status_code == HTTPStatus.OK:
|
| 258 |
ans += response.output.choices[0]['message']['content']
|
| 259 |
-
tk_count += response
|
| 260 |
if response.output.choices[0].get("finish_reason", "") == "length":
|
| 261 |
if is_chinese([ans]):
|
| 262 |
ans += LENGTH_NOTIFICATION_CN
|
|
@@ -292,7 +296,7 @@ class QWenChat(Base):
|
|
| 292 |
for resp in response:
|
| 293 |
if resp.status_code == HTTPStatus.OK:
|
| 294 |
ans = resp.output.choices[0]['message']['content']
|
| 295 |
-
tk_count = resp
|
| 296 |
if resp.output.choices[0].get("finish_reason", "") == "length":
|
| 297 |
if is_chinese(ans):
|
| 298 |
ans += LENGTH_NOTIFICATION_CN
|
|
@@ -334,7 +338,7 @@ class ZhipuChat(Base):
|
|
| 334 |
ans += LENGTH_NOTIFICATION_CN
|
| 335 |
else:
|
| 336 |
ans += LENGTH_NOTIFICATION_EN
|
| 337 |
-
return ans, response
|
| 338 |
except Exception as e:
|
| 339 |
return "**ERROR**: " + str(e), 0
|
| 340 |
|
|
@@ -364,9 +368,9 @@ class ZhipuChat(Base):
|
|
| 364 |
ans += LENGTH_NOTIFICATION_CN
|
| 365 |
else:
|
| 366 |
ans += LENGTH_NOTIFICATION_EN
|
| 367 |
-
tk_count = resp
|
| 368 |
if resp.choices[0].finish_reason == "stop":
|
| 369 |
-
tk_count = resp
|
| 370 |
yield ans
|
| 371 |
except Exception as e:
|
| 372 |
yield ans + "\n**ERROR**: " + str(e)
|
|
@@ -569,7 +573,7 @@ class MiniMaxChat(Base):
|
|
| 569 |
ans += LENGTH_NOTIFICATION_CN
|
| 570 |
else:
|
| 571 |
ans += LENGTH_NOTIFICATION_EN
|
| 572 |
-
return ans, response
|
| 573 |
except Exception as e:
|
| 574 |
return "**ERROR**: " + str(e), 0
|
| 575 |
|
|
@@ -603,11 +607,11 @@ class MiniMaxChat(Base):
|
|
| 603 |
if "choices" in resp and "delta" in resp["choices"][0]:
|
| 604 |
text = resp["choices"][0]["delta"]["content"]
|
| 605 |
ans += text
|
| 606 |
-
|
| 607 |
-
|
| 608 |
-
|
| 609 |
-
|
| 610 |
-
|
| 611 |
yield ans
|
| 612 |
|
| 613 |
except Exception as e:
|
|
@@ -640,7 +644,7 @@ class MistralChat(Base):
|
|
| 640 |
ans += LENGTH_NOTIFICATION_CN
|
| 641 |
else:
|
| 642 |
ans += LENGTH_NOTIFICATION_EN
|
| 643 |
-
return ans, response
|
| 644 |
except openai.APIError as e:
|
| 645 |
return "**ERROR**: " + str(e), 0
|
| 646 |
|
|
@@ -838,7 +842,7 @@ class GeminiChat(Base):
|
|
| 838 |
yield 0
|
| 839 |
|
| 840 |
|
| 841 |
-
class GroqChat:
|
| 842 |
def __init__(self, key, model_name, base_url=''):
|
| 843 |
from groq import Groq
|
| 844 |
self.client = Groq(api_key=key)
|
|
@@ -863,7 +867,7 @@ class GroqChat:
|
|
| 863 |
ans += LENGTH_NOTIFICATION_CN
|
| 864 |
else:
|
| 865 |
ans += LENGTH_NOTIFICATION_EN
|
| 866 |
-
return ans, response
|
| 867 |
except Exception as e:
|
| 868 |
return ans + "\n**ERROR**: " + str(e), 0
|
| 869 |
|
|
@@ -1255,7 +1259,7 @@ class BaiduYiyanChat(Base):
|
|
| 1255 |
**gen_conf
|
| 1256 |
).body
|
| 1257 |
ans = response['result']
|
| 1258 |
-
return ans, response
|
| 1259 |
|
| 1260 |
except Exception as e:
|
| 1261 |
return ans + "\n**ERROR**: " + str(e), 0
|
|
@@ -1283,7 +1287,7 @@ class BaiduYiyanChat(Base):
|
|
| 1283 |
for resp in response:
|
| 1284 |
resp = resp.body
|
| 1285 |
ans += resp['result']
|
| 1286 |
-
total_tokens = resp
|
| 1287 |
|
| 1288 |
yield ans
|
| 1289 |
|
|
|
|
| 53 |
ans += LENGTH_NOTIFICATION_CN
|
| 54 |
else:
|
| 55 |
ans += LENGTH_NOTIFICATION_EN
|
| 56 |
+
return ans, self.total_token_count(response)
|
| 57 |
except openai.APIError as e:
|
| 58 |
return "**ERROR**: " + str(e), 0
|
| 59 |
|
|
|
|
| 75 |
resp.choices[0].delta.content = ""
|
| 76 |
ans += resp.choices[0].delta.content
|
| 77 |
|
| 78 |
+
tol = self.total_token_count(resp)
|
| 79 |
+
if not tol:
|
| 80 |
+
total_tokens += num_tokens_from_string(resp.choices[0].delta.content)
|
|
|
|
|
|
|
|
|
|
|
|
|
| 81 |
else:
|
| 82 |
+
total_tokens = tol
|
| 83 |
|
| 84 |
if resp.choices[0].finish_reason == "length":
|
| 85 |
if is_chinese(ans):
|
|
|
|
| 93 |
|
| 94 |
yield total_tokens
|
| 95 |
|
| 96 |
+
def total_token_count(self, resp):
|
| 97 |
+
try:
|
| 98 |
+
return resp.usage.total_tokens
|
| 99 |
+
except Exception:
|
| 100 |
+
pass
|
| 101 |
+
try:
|
| 102 |
+
return resp["usage"]["total_tokens"]
|
| 103 |
+
except Exception:
|
| 104 |
+
pass
|
| 105 |
+
return 0
|
| 106 |
+
|
| 107 |
|
| 108 |
class GptTurbo(Base):
|
| 109 |
def __init__(self, key, model_name="gpt-3.5-turbo", base_url="https://api.openai.com/v1"):
|
|
|
|
| 189 |
ans += LENGTH_NOTIFICATION_CN
|
| 190 |
else:
|
| 191 |
ans += LENGTH_NOTIFICATION_EN
|
| 192 |
+
return ans, self.total_token_count(response)
|
| 193 |
except openai.APIError as e:
|
| 194 |
return "**ERROR**: " + str(e), 0
|
| 195 |
|
|
|
|
| 219 |
if not resp.choices[0].delta.content:
|
| 220 |
resp.choices[0].delta.content = ""
|
| 221 |
ans += resp.choices[0].delta.content
|
| 222 |
+
tol = self.total_token_count(resp)
|
| 223 |
+
if not tol:
|
| 224 |
+
total_tokens += num_tokens_from_string(resp.choices[0].delta.content)
|
| 225 |
+
else:
|
| 226 |
+
total_tokens = tol
|
|
|
|
|
|
|
|
|
|
| 227 |
if resp.choices[0].finish_reason == "length":
|
| 228 |
if is_chinese([ans]):
|
| 229 |
ans += LENGTH_NOTIFICATION_CN
|
|
|
|
| 260 |
tk_count = 0
|
| 261 |
if response.status_code == HTTPStatus.OK:
|
| 262 |
ans += response.output.choices[0]['message']['content']
|
| 263 |
+
tk_count += self.total_token_count(response)
|
| 264 |
if response.output.choices[0].get("finish_reason", "") == "length":
|
| 265 |
if is_chinese([ans]):
|
| 266 |
ans += LENGTH_NOTIFICATION_CN
|
|
|
|
| 296 |
for resp in response:
|
| 297 |
if resp.status_code == HTTPStatus.OK:
|
| 298 |
ans = resp.output.choices[0]['message']['content']
|
| 299 |
+
tk_count = self.total_token_count(resp)
|
| 300 |
if resp.output.choices[0].get("finish_reason", "") == "length":
|
| 301 |
if is_chinese(ans):
|
| 302 |
ans += LENGTH_NOTIFICATION_CN
|
|
|
|
| 338 |
ans += LENGTH_NOTIFICATION_CN
|
| 339 |
else:
|
| 340 |
ans += LENGTH_NOTIFICATION_EN
|
| 341 |
+
return ans, self.total_token_count(response)
|
| 342 |
except Exception as e:
|
| 343 |
return "**ERROR**: " + str(e), 0
|
| 344 |
|
|
|
|
| 368 |
ans += LENGTH_NOTIFICATION_CN
|
| 369 |
else:
|
| 370 |
ans += LENGTH_NOTIFICATION_EN
|
| 371 |
+
tk_count = self.total_token_count(resp)
|
| 372 |
if resp.choices[0].finish_reason == "stop":
|
| 373 |
+
tk_count = self.total_token_count(resp)
|
| 374 |
yield ans
|
| 375 |
except Exception as e:
|
| 376 |
yield ans + "\n**ERROR**: " + str(e)
|
|
|
|
| 573 |
ans += LENGTH_NOTIFICATION_CN
|
| 574 |
else:
|
| 575 |
ans += LENGTH_NOTIFICATION_EN
|
| 576 |
+
return ans, self.total_token_count(response)
|
| 577 |
except Exception as e:
|
| 578 |
return "**ERROR**: " + str(e), 0
|
| 579 |
|
|
|
|
| 607 |
if "choices" in resp and "delta" in resp["choices"][0]:
|
| 608 |
text = resp["choices"][0]["delta"]["content"]
|
| 609 |
ans += text
|
| 610 |
+
tol = self.total_token_count(resp)
|
| 611 |
+
if not tol:
|
| 612 |
+
total_tokens += num_tokens_from_string(text)
|
| 613 |
+
else:
|
| 614 |
+
total_tokens = tol
|
| 615 |
yield ans
|
| 616 |
|
| 617 |
except Exception as e:
|
|
|
|
| 644 |
ans += LENGTH_NOTIFICATION_CN
|
| 645 |
else:
|
| 646 |
ans += LENGTH_NOTIFICATION_EN
|
| 647 |
+
return ans, self.total_token_count(response)
|
| 648 |
except openai.APIError as e:
|
| 649 |
return "**ERROR**: " + str(e), 0
|
| 650 |
|
|
|
|
| 842 |
yield 0
|
| 843 |
|
| 844 |
|
| 845 |
+
class GroqChat(Base):
|
| 846 |
def __init__(self, key, model_name, base_url=''):
|
| 847 |
from groq import Groq
|
| 848 |
self.client = Groq(api_key=key)
|
|
|
|
| 867 |
ans += LENGTH_NOTIFICATION_CN
|
| 868 |
else:
|
| 869 |
ans += LENGTH_NOTIFICATION_EN
|
| 870 |
+
return ans, self.total_token_count(response)
|
| 871 |
except Exception as e:
|
| 872 |
return ans + "\n**ERROR**: " + str(e), 0
|
| 873 |
|
|
|
|
| 1259 |
**gen_conf
|
| 1260 |
).body
|
| 1261 |
ans = response['result']
|
| 1262 |
+
return ans, self.total_token_count(response)
|
| 1263 |
|
| 1264 |
except Exception as e:
|
| 1265 |
return ans + "\n**ERROR**: " + str(e), 0
|
|
|
|
| 1287 |
for resp in response:
|
| 1288 |
resp = resp.body
|
| 1289 |
ans += resp['result']
|
| 1290 |
+
total_tokens = self.total_token_count(resp)
|
| 1291 |
|
| 1292 |
yield ans
|
| 1293 |
|
rag/llm/embedding_model.py
CHANGED
|
@@ -44,11 +44,23 @@ class Base(ABC):
|
|
| 44 |
def encode_queries(self, text: str):
|
| 45 |
raise NotImplementedError("Please implement encode method!")
|
| 46 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 47 |
|
| 48 |
class DefaultEmbedding(Base):
|
| 49 |
_model = None
|
| 50 |
_model_name = ""
|
| 51 |
_model_lock = threading.Lock()
|
|
|
|
| 52 |
def __init__(self, key, model_name, **kwargs):
|
| 53 |
"""
|
| 54 |
If you have trouble downloading HuggingFace models, -_^ this might help!!
|
|
@@ -115,13 +127,13 @@ class OpenAIEmbed(Base):
|
|
| 115 |
res = self.client.embeddings.create(input=texts[i:i + batch_size],
|
| 116 |
model=self.model_name)
|
| 117 |
ress.extend([d.embedding for d in res.data])
|
| 118 |
-
total_tokens += res
|
| 119 |
return np.array(ress), total_tokens
|
| 120 |
|
| 121 |
def encode_queries(self, text):
|
| 122 |
res = self.client.embeddings.create(input=[truncate(text, 8191)],
|
| 123 |
model=self.model_name)
|
| 124 |
-
return np.array(res.data[0].embedding), res
|
| 125 |
|
| 126 |
|
| 127 |
class LocalAIEmbed(Base):
|
|
@@ -188,7 +200,7 @@ class QWenEmbed(Base):
|
|
| 188 |
for e in resp["output"]["embeddings"]:
|
| 189 |
embds[e["text_index"]] = e["embedding"]
|
| 190 |
res.extend(embds)
|
| 191 |
-
token_count += resp
|
| 192 |
return np.array(res), token_count
|
| 193 |
except Exception as e:
|
| 194 |
raise Exception("Account abnormal. Please ensure it's on good standing to use QWen's "+self.model_name)
|
|
@@ -203,7 +215,7 @@ class QWenEmbed(Base):
|
|
| 203 |
text_type="query"
|
| 204 |
)
|
| 205 |
return np.array(resp["output"]["embeddings"][0]
|
| 206 |
-
["embedding"]), resp
|
| 207 |
except Exception:
|
| 208 |
raise Exception("Account abnormal. Please ensure it's on good standing to use QWen's "+self.model_name)
|
| 209 |
return np.array([]), 0
|
|
@@ -229,13 +241,13 @@ class ZhipuEmbed(Base):
|
|
| 229 |
res = self.client.embeddings.create(input=txt,
|
| 230 |
model=self.model_name)
|
| 231 |
arr.append(res.data[0].embedding)
|
| 232 |
-
tks_num += res
|
| 233 |
return np.array(arr), tks_num
|
| 234 |
|
| 235 |
def encode_queries(self, text):
|
| 236 |
res = self.client.embeddings.create(input=text,
|
| 237 |
model=self.model_name)
|
| 238 |
-
return np.array(res.data[0].embedding), res
|
| 239 |
|
| 240 |
|
| 241 |
class OllamaEmbed(Base):
|
|
@@ -318,13 +330,13 @@ class XinferenceEmbed(Base):
|
|
| 318 |
for i in range(0, len(texts), batch_size):
|
| 319 |
res = self.client.embeddings.create(input=texts[i:i + batch_size], model=self.model_name)
|
| 320 |
ress.extend([d.embedding for d in res.data])
|
| 321 |
-
total_tokens += res
|
| 322 |
return np.array(ress), total_tokens
|
| 323 |
|
| 324 |
def encode_queries(self, text):
|
| 325 |
res = self.client.embeddings.create(input=[text],
|
| 326 |
model=self.model_name)
|
| 327 |
-
return np.array(res.data[0].embedding), res
|
| 328 |
|
| 329 |
|
| 330 |
class YoudaoEmbed(Base):
|
|
@@ -383,7 +395,7 @@ class JinaEmbed(Base):
|
|
| 383 |
}
|
| 384 |
res = requests.post(self.base_url, headers=self.headers, json=data).json()
|
| 385 |
ress.extend([d["embedding"] for d in res["data"]])
|
| 386 |
-
token_count += res
|
| 387 |
return np.array(ress), token_count
|
| 388 |
|
| 389 |
def encode_queries(self, text):
|
|
@@ -447,13 +459,13 @@ class MistralEmbed(Base):
|
|
| 447 |
res = self.client.embeddings(input=texts[i:i + batch_size],
|
| 448 |
model=self.model_name)
|
| 449 |
ress.extend([d.embedding for d in res.data])
|
| 450 |
-
token_count += res
|
| 451 |
return np.array(ress), token_count
|
| 452 |
|
| 453 |
def encode_queries(self, text):
|
| 454 |
res = self.client.embeddings(input=[truncate(text, 8196)],
|
| 455 |
model=self.model_name)
|
| 456 |
-
return np.array(res.data[0].embedding), res
|
| 457 |
|
| 458 |
|
| 459 |
class BedrockEmbed(Base):
|
|
@@ -565,7 +577,7 @@ class NvidiaEmbed(Base):
|
|
| 565 |
}
|
| 566 |
res = requests.post(self.base_url, headers=self.headers, json=payload).json()
|
| 567 |
ress.extend([d["embedding"] for d in res["data"]])
|
| 568 |
-
token_count += res
|
| 569 |
return np.array(ress), token_count
|
| 570 |
|
| 571 |
def encode_queries(self, text):
|
|
@@ -677,7 +689,7 @@ class SILICONFLOWEmbed(Base):
|
|
| 677 |
if "data" not in res or not isinstance(res["data"], list) or len(res["data"]) != len(texts_batch):
|
| 678 |
raise ValueError(f"SILICONFLOWEmbed.encode got invalid response from {self.base_url}")
|
| 679 |
ress.extend([d["embedding"] for d in res["data"]])
|
| 680 |
-
token_count += res
|
| 681 |
return np.array(ress), token_count
|
| 682 |
|
| 683 |
def encode_queries(self, text):
|
|
@@ -689,7 +701,7 @@ class SILICONFLOWEmbed(Base):
|
|
| 689 |
res = requests.post(self.base_url, json=payload, headers=self.headers).json()
|
| 690 |
if "data" not in res or not isinstance(res["data"], list) or len(res["data"])!= 1:
|
| 691 |
raise ValueError(f"SILICONFLOWEmbed.encode_queries got invalid response from {self.base_url}")
|
| 692 |
-
return np.array(res["data"][0]["embedding"]), res
|
| 693 |
|
| 694 |
|
| 695 |
class ReplicateEmbed(Base):
|
|
@@ -727,14 +739,14 @@ class BaiduYiyanEmbed(Base):
|
|
| 727 |
res = self.client.do(model=self.model_name, texts=texts).body
|
| 728 |
return (
|
| 729 |
np.array([r["embedding"] for r in res["data"]]),
|
| 730 |
-
res
|
| 731 |
)
|
| 732 |
|
| 733 |
def encode_queries(self, text):
|
| 734 |
res = self.client.do(model=self.model_name, texts=[text]).body
|
| 735 |
return (
|
| 736 |
np.array([r["embedding"] for r in res["data"]]),
|
| 737 |
-
res
|
| 738 |
)
|
| 739 |
|
| 740 |
|
|
|
|
| 44 |
def encode_queries(self, text: str):
|
| 45 |
raise NotImplementedError("Please implement encode method!")
|
| 46 |
|
| 47 |
+
def total_token_count(self, resp):
|
| 48 |
+
try:
|
| 49 |
+
return resp.usage.total_tokens
|
| 50 |
+
except Exception:
|
| 51 |
+
pass
|
| 52 |
+
try:
|
| 53 |
+
return resp["usage"]["total_tokens"]
|
| 54 |
+
except Exception:
|
| 55 |
+
pass
|
| 56 |
+
return 0
|
| 57 |
+
|
| 58 |
|
| 59 |
class DefaultEmbedding(Base):
|
| 60 |
_model = None
|
| 61 |
_model_name = ""
|
| 62 |
_model_lock = threading.Lock()
|
| 63 |
+
|
| 64 |
def __init__(self, key, model_name, **kwargs):
|
| 65 |
"""
|
| 66 |
If you have trouble downloading HuggingFace models, -_^ this might help!!
|
|
|
|
| 127 |
res = self.client.embeddings.create(input=texts[i:i + batch_size],
|
| 128 |
model=self.model_name)
|
| 129 |
ress.extend([d.embedding for d in res.data])
|
| 130 |
+
total_tokens += self.total_token_count(res)
|
| 131 |
return np.array(ress), total_tokens
|
| 132 |
|
| 133 |
def encode_queries(self, text):
|
| 134 |
res = self.client.embeddings.create(input=[truncate(text, 8191)],
|
| 135 |
model=self.model_name)
|
| 136 |
+
return np.array(res.data[0].embedding), self.total_token_count(res)
|
| 137 |
|
| 138 |
|
| 139 |
class LocalAIEmbed(Base):
|
|
|
|
| 200 |
for e in resp["output"]["embeddings"]:
|
| 201 |
embds[e["text_index"]] = e["embedding"]
|
| 202 |
res.extend(embds)
|
| 203 |
+
token_count += self.total_token_count(resp)
|
| 204 |
return np.array(res), token_count
|
| 205 |
except Exception as e:
|
| 206 |
raise Exception("Account abnormal. Please ensure it's on good standing to use QWen's "+self.model_name)
|
|
|
|
| 215 |
text_type="query"
|
| 216 |
)
|
| 217 |
return np.array(resp["output"]["embeddings"][0]
|
| 218 |
+
["embedding"]), self.total_token_count(resp)
|
| 219 |
except Exception:
|
| 220 |
raise Exception("Account abnormal. Please ensure it's on good standing to use QWen's "+self.model_name)
|
| 221 |
return np.array([]), 0
|
|
|
|
| 241 |
res = self.client.embeddings.create(input=txt,
|
| 242 |
model=self.model_name)
|
| 243 |
arr.append(res.data[0].embedding)
|
| 244 |
+
tks_num += self.total_token_count(res)
|
| 245 |
return np.array(arr), tks_num
|
| 246 |
|
| 247 |
def encode_queries(self, text):
|
| 248 |
res = self.client.embeddings.create(input=text,
|
| 249 |
model=self.model_name)
|
| 250 |
+
return np.array(res.data[0].embedding), self.total_token_count(res)
|
| 251 |
|
| 252 |
|
| 253 |
class OllamaEmbed(Base):
|
|
|
|
| 330 |
for i in range(0, len(texts), batch_size):
|
| 331 |
res = self.client.embeddings.create(input=texts[i:i + batch_size], model=self.model_name)
|
| 332 |
ress.extend([d.embedding for d in res.data])
|
| 333 |
+
total_tokens += self.total_token_count(res)
|
| 334 |
return np.array(ress), total_tokens
|
| 335 |
|
| 336 |
def encode_queries(self, text):
|
| 337 |
res = self.client.embeddings.create(input=[text],
|
| 338 |
model=self.model_name)
|
| 339 |
+
return np.array(res.data[0].embedding), self.total_token_count(res)
|
| 340 |
|
| 341 |
|
| 342 |
class YoudaoEmbed(Base):
|
|
|
|
| 395 |
}
|
| 396 |
res = requests.post(self.base_url, headers=self.headers, json=data).json()
|
| 397 |
ress.extend([d["embedding"] for d in res["data"]])
|
| 398 |
+
token_count += self.total_token_count(res)
|
| 399 |
return np.array(ress), token_count
|
| 400 |
|
| 401 |
def encode_queries(self, text):
|
|
|
|
| 459 |
res = self.client.embeddings(input=texts[i:i + batch_size],
|
| 460 |
model=self.model_name)
|
| 461 |
ress.extend([d.embedding for d in res.data])
|
| 462 |
+
token_count += self.total_token_count(res)
|
| 463 |
return np.array(ress), token_count
|
| 464 |
|
| 465 |
def encode_queries(self, text):
|
| 466 |
res = self.client.embeddings(input=[truncate(text, 8196)],
|
| 467 |
model=self.model_name)
|
| 468 |
+
return np.array(res.data[0].embedding), self.total_token_count(res)
|
| 469 |
|
| 470 |
|
| 471 |
class BedrockEmbed(Base):
|
|
|
|
| 577 |
}
|
| 578 |
res = requests.post(self.base_url, headers=self.headers, json=payload).json()
|
| 579 |
ress.extend([d["embedding"] for d in res["data"]])
|
| 580 |
+
token_count += self.total_token_count(res)
|
| 581 |
return np.array(ress), token_count
|
| 582 |
|
| 583 |
def encode_queries(self, text):
|
|
|
|
| 689 |
if "data" not in res or not isinstance(res["data"], list) or len(res["data"]) != len(texts_batch):
|
| 690 |
raise ValueError(f"SILICONFLOWEmbed.encode got invalid response from {self.base_url}")
|
| 691 |
ress.extend([d["embedding"] for d in res["data"]])
|
| 692 |
+
token_count += self.total_token_count(res)
|
| 693 |
return np.array(ress), token_count
|
| 694 |
|
| 695 |
def encode_queries(self, text):
|
|
|
|
| 701 |
res = requests.post(self.base_url, json=payload, headers=self.headers).json()
|
| 702 |
if "data" not in res or not isinstance(res["data"], list) or len(res["data"])!= 1:
|
| 703 |
raise ValueError(f"SILICONFLOWEmbed.encode_queries got invalid response from {self.base_url}")
|
| 704 |
+
return np.array(res["data"][0]["embedding"]), self.total_token_count(res)
|
| 705 |
|
| 706 |
|
| 707 |
class ReplicateEmbed(Base):
|
|
|
|
| 739 |
res = self.client.do(model=self.model_name, texts=texts).body
|
| 740 |
return (
|
| 741 |
np.array([r["embedding"] for r in res["data"]]),
|
| 742 |
+
self.total_token_count(res),
|
| 743 |
)
|
| 744 |
|
| 745 |
def encode_queries(self, text):
|
| 746 |
res = self.client.do(model=self.model_name, texts=[text]).body
|
| 747 |
return (
|
| 748 |
np.array([r["embedding"] for r in res["data"]]),
|
| 749 |
+
self.total_token_count(res),
|
| 750 |
)
|
| 751 |
|
| 752 |
|
rag/llm/rerank_model.py
CHANGED
|
@@ -42,6 +42,17 @@ class Base(ABC):
|
|
| 42 |
def similarity(self, query: str, texts: list):
|
| 43 |
raise NotImplementedError("Please implement encode method!")
|
| 44 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 45 |
|
| 46 |
class DefaultRerank(Base):
|
| 47 |
_model = None
|
|
@@ -115,7 +126,7 @@ class JinaRerank(Base):
|
|
| 115 |
rank = np.zeros(len(texts), dtype=float)
|
| 116 |
for d in res["results"]:
|
| 117 |
rank[d["index"]] = d["relevance_score"]
|
| 118 |
-
return rank, res
|
| 119 |
|
| 120 |
|
| 121 |
class YoudaoRerank(DefaultRerank):
|
|
@@ -417,7 +428,7 @@ class BaiduYiyanRerank(Base):
|
|
| 417 |
rank = np.zeros(len(texts), dtype=float)
|
| 418 |
for d in res["results"]:
|
| 419 |
rank[d["index"]] = d["relevance_score"]
|
| 420 |
-
return rank, res
|
| 421 |
|
| 422 |
|
| 423 |
class VoyageRerank(Base):
|
|
|
|
| 42 |
def similarity(self, query: str, texts: list):
|
| 43 |
raise NotImplementedError("Please implement encode method!")
|
| 44 |
|
| 45 |
+
def total_token_count(self, resp):
|
| 46 |
+
try:
|
| 47 |
+
return resp.usage.total_tokens
|
| 48 |
+
except Exception:
|
| 49 |
+
pass
|
| 50 |
+
try:
|
| 51 |
+
return resp["usage"]["total_tokens"]
|
| 52 |
+
except Exception:
|
| 53 |
+
pass
|
| 54 |
+
return 0
|
| 55 |
+
|
| 56 |
|
| 57 |
class DefaultRerank(Base):
|
| 58 |
_model = None
|
|
|
|
| 126 |
rank = np.zeros(len(texts), dtype=float)
|
| 127 |
for d in res["results"]:
|
| 128 |
rank[d["index"]] = d["relevance_score"]
|
| 129 |
+
return rank, self.total_token_count(res)
|
| 130 |
|
| 131 |
|
| 132 |
class YoudaoRerank(DefaultRerank):
|
|
|
|
| 428 |
rank = np.zeros(len(texts), dtype=float)
|
| 429 |
for d in res["results"]:
|
| 430 |
rank[d["index"]] = d["relevance_score"]
|
| 431 |
+
return rank, self.total_token_count(res)
|
| 432 |
|
| 433 |
|
| 434 |
class VoyageRerank(Base):
|