Spaces:
Sleeping
Sleeping
fahmiaziz98
commited on
Commit
Β·
073edba
1
Parent(s):
76d149a
restapi
Browse files- app.py +514 -5
- requirements.txt +6 -2
app.py
CHANGED
|
@@ -1,7 +1,516 @@
|
|
| 1 |
-
from fastapi import FastAPI
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 2 |
|
| 3 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 4 |
|
| 5 |
-
@app.get("/")
|
| 6 |
-
def greet_json():
|
| 7 |
-
return {"Hello": "World!"}
|
|
|
|
| 1 |
+
from fastapi import FastAPI, HTTPException
|
| 2 |
+
from pydantic import BaseModel, Field
|
| 3 |
+
from typing import List, Optional, Dict, Any
|
| 4 |
+
from loguru import logger
|
| 5 |
+
import time
|
| 6 |
+
import torch
|
| 7 |
+
from contextlib import asynccontextmanager
|
| 8 |
|
| 9 |
+
from sentence_transformers import CrossEncoder
|
| 10 |
+
from transformers import AutoTokenizer, AutoModelForCausalLM
|
| 11 |
+
|
| 12 |
+
|
| 13 |
+
# -------------------------
|
| 14 |
+
# Request/Response Models
|
| 15 |
+
# -------------------------
|
| 16 |
+
|
| 17 |
+
class RerankRequest(BaseModel):
|
| 18 |
+
"""
|
| 19 |
+
Request model for document reranking.
|
| 20 |
+
|
| 21 |
+
Attributes:
|
| 22 |
+
query: The search query
|
| 23 |
+
documents: List of documents to rerank
|
| 24 |
+
model_id: Identifier of the reranking model to use
|
| 25 |
+
instruction: Optional instruction for instruction-based models
|
| 26 |
+
top_k: Maximum number of documents to return (optional)
|
| 27 |
+
"""
|
| 28 |
+
query: str = Field(..., description="Search query text")
|
| 29 |
+
documents: List[str] = Field(..., min_items=1, description="List of documents to rerank")
|
| 30 |
+
model_id: str = Field(..., description="Model identifier for reranking")
|
| 31 |
+
instruction: Optional[str] = Field(None, description="Optional instruction for reranking task")
|
| 32 |
+
top_k: Optional[int] = Field(None, description="Maximum number of results to return")
|
| 33 |
+
|
| 34 |
+
|
| 35 |
+
class RerankResult(BaseModel):
|
| 36 |
+
"""
|
| 37 |
+
Single reranking result.
|
| 38 |
+
|
| 39 |
+
Attributes:
|
| 40 |
+
text: The document text
|
| 41 |
+
score: Relevance score from the reranking model
|
| 42 |
+
index: Original index of the document in input list
|
| 43 |
+
"""
|
| 44 |
+
text: str
|
| 45 |
+
score: float
|
| 46 |
+
index: int
|
| 47 |
+
|
| 48 |
+
|
| 49 |
+
class RerankResponse(BaseModel):
|
| 50 |
+
"""
|
| 51 |
+
Response model for document reranking.
|
| 52 |
+
|
| 53 |
+
Attributes:
|
| 54 |
+
results: List of reranked documents with scores
|
| 55 |
+
query: The original search query
|
| 56 |
+
model_id: Identifier of the model used
|
| 57 |
+
processing_time: Time taken to process the request
|
| 58 |
+
total_documents: Total number of input documents
|
| 59 |
+
returned_documents: Number of documents returned
|
| 60 |
+
"""
|
| 61 |
+
results: List[RerankResult]
|
| 62 |
+
query: str
|
| 63 |
+
model_id: str
|
| 64 |
+
processing_time: float
|
| 65 |
+
total_documents: int
|
| 66 |
+
returned_documents: int
|
| 67 |
+
|
| 68 |
+
|
| 69 |
+
# -------------------------
|
| 70 |
+
# Model Management
|
| 71 |
+
# -------------------------
|
| 72 |
+
|
| 73 |
+
class RerankerModel:
|
| 74 |
+
"""Base class for reranking models."""
|
| 75 |
+
|
| 76 |
+
def __init__(self, model_id: str, model_name: str, model_type: str):
|
| 77 |
+
self.model_id = model_id
|
| 78 |
+
self.model_name = model_name
|
| 79 |
+
self.model_type = model_type
|
| 80 |
+
self.model = None
|
| 81 |
+
self.tokenizer = None
|
| 82 |
+
self.loaded = False
|
| 83 |
+
|
| 84 |
+
def load(self):
|
| 85 |
+
"""Load the model. To be implemented by subclasses."""
|
| 86 |
+
raise NotImplementedError
|
| 87 |
+
|
| 88 |
+
def rerank(self, query: str, documents: List[str], instruction: Optional[str] = None) -> List[float]:
|
| 89 |
+
"""Rerank documents. To be implemented by subclasses."""
|
| 90 |
+
raise NotImplementedError
|
| 91 |
+
|
| 92 |
+
|
| 93 |
+
class SentenceTransformersReranker(RerankerModel):
|
| 94 |
+
"""Reranker using sentence-transformers CrossEncoder."""
|
| 95 |
+
|
| 96 |
+
def load(self):
|
| 97 |
+
"""Load sentence-transformers CrossEncoder model."""
|
| 98 |
+
try:
|
| 99 |
+
logger.info(f"Loading SentenceTransformers model: {self.model_name}")
|
| 100 |
+
self.model = CrossEncoder(
|
| 101 |
+
self.model_name,
|
| 102 |
+
model_kwargs={"torch_dtype": "auto"},
|
| 103 |
+
trust_remote_code=True
|
| 104 |
+
)
|
| 105 |
+
self.loaded = True
|
| 106 |
+
logger.success(f"Successfully loaded {self.model_id}")
|
| 107 |
+
except Exception as e:
|
| 108 |
+
logger.error(f"Failed to load {self.model_id}: {e}")
|
| 109 |
+
raise
|
| 110 |
+
|
| 111 |
+
def rerank(self, query: str, documents: List[str], instruction: Optional[str] = None) -> List[float]:
|
| 112 |
+
"""Rerank documents using CrossEncoder."""
|
| 113 |
+
if not self.loaded:
|
| 114 |
+
raise RuntimeError(f"Model {self.model_id} not loaded")
|
| 115 |
+
|
| 116 |
+
try:
|
| 117 |
+
# For sentence-transformers, we can use the rank method directly
|
| 118 |
+
rankings = self.model.rank(query, documents, convert_to_tensor=True)
|
| 119 |
+
|
| 120 |
+
# Extract scores and maintain original order
|
| 121 |
+
scores = [0.0] * len(documents)
|
| 122 |
+
for ranking in rankings:
|
| 123 |
+
scores[ranking['corpus_id']] = float(ranking['score'])
|
| 124 |
+
|
| 125 |
+
return scores
|
| 126 |
+
|
| 127 |
+
except Exception as e:
|
| 128 |
+
logger.error(f"Reranking failed with {self.model_id}: {e}")
|
| 129 |
+
raise
|
| 130 |
+
|
| 131 |
+
|
| 132 |
+
class QwenReranker(RerankerModel):
|
| 133 |
+
"""Reranker using Qwen3-Reranker model."""
|
| 134 |
+
|
| 135 |
+
def load(self):
|
| 136 |
+
"""Load Qwen reranker model."""
|
| 137 |
+
try:
|
| 138 |
+
logger.info(f"Loading Qwen model: {self.model_name}")
|
| 139 |
+
|
| 140 |
+
self.tokenizer = AutoTokenizer.from_pretrained(
|
| 141 |
+
self.model_name,
|
| 142 |
+
padding_side='left'
|
| 143 |
+
)
|
| 144 |
+
self.model = AutoModelForCausalLM.from_pretrained(
|
| 145 |
+
self.model_name
|
| 146 |
+
).eval()
|
| 147 |
+
|
| 148 |
+
# Set up Qwen-specific tokens
|
| 149 |
+
self.token_false_id = self.tokenizer.convert_tokens_to_ids("no")
|
| 150 |
+
self.token_true_id = self.tokenizer.convert_tokens_to_ids("yes")
|
| 151 |
+
self.max_length = 8192
|
| 152 |
+
|
| 153 |
+
# Set up prompt templates
|
| 154 |
+
self.prefix = "<|im_start|>system\nJudge whether the Document meets the requirements based on the Query and the Instruct provided. Note that the answer can only be \"yes\" or \"no\".<|im_end|>\n<|im_start|>user\n"
|
| 155 |
+
self.suffix = "<|im_end|>\n<|im_start|>assistant\n<think>\n\n</think>\n\n"
|
| 156 |
+
self.prefix_tokens = self.tokenizer.encode(self.prefix, add_special_tokens=False)
|
| 157 |
+
self.suffix_tokens = self.tokenizer.encode(self.suffix, add_special_tokens=False)
|
| 158 |
+
|
| 159 |
+
self.loaded = True
|
| 160 |
+
logger.success(f"Successfully loaded {self.model_id}")
|
| 161 |
+
|
| 162 |
+
except Exception as e:
|
| 163 |
+
logger.error(f"Failed to load {self.model_id}: {e}")
|
| 164 |
+
raise
|
| 165 |
+
|
| 166 |
+
def _format_instruction(self, instruction: str, query: str, doc: str) -> str:
|
| 167 |
+
"""Format instruction for Qwen model."""
|
| 168 |
+
if instruction is None:
|
| 169 |
+
instruction = 'Given a web search query, retrieve relevant passages that answer the query'
|
| 170 |
+
|
| 171 |
+
return "<Instruct>: {instruction}\n<Query>: {query}\n<Document>: {doc}".format(
|
| 172 |
+
instruction=instruction, query=query, doc=doc
|
| 173 |
+
)
|
| 174 |
+
|
| 175 |
+
def _process_inputs(self, pairs: List[str]):
|
| 176 |
+
"""Process input pairs for Qwen model."""
|
| 177 |
+
inputs = self.tokenizer(
|
| 178 |
+
pairs,
|
| 179 |
+
padding=False,
|
| 180 |
+
truncation='longest_first',
|
| 181 |
+
return_attention_mask=False,
|
| 182 |
+
max_length=self.max_length - len(self.prefix_tokens) - len(self.suffix_tokens)
|
| 183 |
+
)
|
| 184 |
+
|
| 185 |
+
for i, ele in enumerate(inputs['input_ids']):
|
| 186 |
+
inputs['input_ids'][i] = self.prefix_tokens + ele + self.suffix_tokens
|
| 187 |
+
|
| 188 |
+
inputs = self.tokenizer.pad(
|
| 189 |
+
inputs,
|
| 190 |
+
padding=True,
|
| 191 |
+
return_tensors="pt",
|
| 192 |
+
max_length=self.max_length
|
| 193 |
+
)
|
| 194 |
+
|
| 195 |
+
for key in inputs:
|
| 196 |
+
inputs[key] = inputs[key].to(self.model.device)
|
| 197 |
+
|
| 198 |
+
return inputs
|
| 199 |
+
|
| 200 |
+
@torch.no_grad()
|
| 201 |
+
def _compute_logits(self, inputs):
|
| 202 |
+
"""Compute relevance scores from model logits."""
|
| 203 |
+
batch_scores = self.model(**inputs).logits[:, -1, :]
|
| 204 |
+
true_vector = batch_scores[:, self.token_true_id]
|
| 205 |
+
false_vector = batch_scores[:, self.token_false_id]
|
| 206 |
+
batch_scores = torch.stack([false_vector, true_vector], dim=1)
|
| 207 |
+
batch_scores = torch.nn.functional.log_softmax(batch_scores, dim=1)
|
| 208 |
+
scores = batch_scores[:, 1].exp().tolist()
|
| 209 |
+
return scores
|
| 210 |
+
|
| 211 |
+
def rerank(self, query: str, documents: List[str], instruction: Optional[str] = None) -> List[float]:
|
| 212 |
+
"""Rerank documents using Qwen model."""
|
| 213 |
+
if not self.loaded:
|
| 214 |
+
raise RuntimeError(f"Model {self.model_id} not loaded")
|
| 215 |
+
|
| 216 |
+
try:
|
| 217 |
+
# Format instruction pairs
|
| 218 |
+
pairs = [
|
| 219 |
+
self._format_instruction(instruction, query, doc)
|
| 220 |
+
for doc in documents
|
| 221 |
+
]
|
| 222 |
+
|
| 223 |
+
# Process inputs
|
| 224 |
+
inputs = self._process_inputs(pairs)
|
| 225 |
+
|
| 226 |
+
# Compute scores
|
| 227 |
+
scores = self._compute_logits(inputs)
|
| 228 |
+
|
| 229 |
+
return scores
|
| 230 |
+
|
| 231 |
+
except Exception as e:
|
| 232 |
+
logger.error(f"Reranking failed with {self.model_id}: {e}")
|
| 233 |
+
raise
|
| 234 |
+
|
| 235 |
+
|
| 236 |
+
class ModelManager:
|
| 237 |
+
"""Manager for reranking models with preloading."""
|
| 238 |
+
|
| 239 |
+
def __init__(self):
|
| 240 |
+
self.models: Dict[str, RerankerModel] = {}
|
| 241 |
+
self.model_configs = {
|
| 242 |
+
"jina-reranker-v2": {
|
| 243 |
+
"model_name": "jinaai/jina-reranker-v2-base-multilingual",
|
| 244 |
+
"model_type": "sentence_transformers",
|
| 245 |
+
"description": "Multilingual reranker from Jina AI"
|
| 246 |
+
},
|
| 247 |
+
"bge-reranker-v2": {
|
| 248 |
+
"model_name": "BAAI/bge-reranker-v2-m3",
|
| 249 |
+
"model_type": "sentence_transformers",
|
| 250 |
+
"description": "BGE multilingual reranker"
|
| 251 |
+
},
|
| 252 |
+
"qwen3-reranker": {
|
| 253 |
+
"model_name": "Qwen/Qwen3-Reranker-0.6B",
|
| 254 |
+
"model_type": "qwen",
|
| 255 |
+
"description": "Qwen3 instruction-based reranker"
|
| 256 |
+
}
|
| 257 |
+
}
|
| 258 |
+
|
| 259 |
+
async def preload_all_models(self):
|
| 260 |
+
"""Preload all configured models."""
|
| 261 |
+
logger.info(f"Starting preload of {len(self.model_configs)} reranking models...")
|
| 262 |
+
|
| 263 |
+
for model_id, config in self.model_configs.items():
|
| 264 |
+
try:
|
| 265 |
+
logger.info(f"Loading {model_id}...")
|
| 266 |
+
|
| 267 |
+
if config["model_type"] == "sentence_transformers":
|
| 268 |
+
model = SentenceTransformersReranker(
|
| 269 |
+
model_id=model_id,
|
| 270 |
+
model_name=config["model_name"],
|
| 271 |
+
model_type=config["model_type"]
|
| 272 |
+
)
|
| 273 |
+
elif config["model_type"] == "qwen":
|
| 274 |
+
model = QwenReranker(
|
| 275 |
+
model_id=model_id,
|
| 276 |
+
model_name=config["model_name"],
|
| 277 |
+
model_type=config["model_type"]
|
| 278 |
+
)
|
| 279 |
+
else:
|
| 280 |
+
logger.error(f"Unknown model type: {config['model_type']}")
|
| 281 |
+
continue
|
| 282 |
+
|
| 283 |
+
model.load()
|
| 284 |
+
self.models[model_id] = model
|
| 285 |
+
logger.success(f"Successfully preloaded {model_id}")
|
| 286 |
+
|
| 287 |
+
except Exception as e:
|
| 288 |
+
logger.error(f"Failed to preload {model_id}: {e}")
|
| 289 |
+
|
| 290 |
+
loaded_count = len([m for m in self.models.values() if m.loaded])
|
| 291 |
+
logger.success(f"Preloaded {loaded_count}/{len(self.model_configs)} models successfully")
|
| 292 |
+
|
| 293 |
+
def get_model(self, model_id: str) -> RerankerModel:
|
| 294 |
+
"""Get a loaded model by ID."""
|
| 295 |
+
if model_id not in self.models:
|
| 296 |
+
raise ValueError(f"Model {model_id} not found")
|
| 297 |
+
|
| 298 |
+
model = self.models[model_id]
|
| 299 |
+
if not model.loaded:
|
| 300 |
+
raise ValueError(f"Model {model_id} not loaded")
|
| 301 |
+
|
| 302 |
+
return model
|
| 303 |
+
|
| 304 |
+
def list_models(self) -> List[Dict[str, Any]]:
|
| 305 |
+
"""List all available models with their status."""
|
| 306 |
+
models_info = []
|
| 307 |
+
for model_id, config in self.model_configs.items():
|
| 308 |
+
model = self.models.get(model_id)
|
| 309 |
+
info = {
|
| 310 |
+
"id": model_id,
|
| 311 |
+
"name": config["model_name"],
|
| 312 |
+
"type": config["model_type"],
|
| 313 |
+
"description": config["description"],
|
| 314 |
+
"loaded": model.loaded if model else False
|
| 315 |
+
}
|
| 316 |
+
models_info.append(info)
|
| 317 |
+
return models_info
|
| 318 |
+
|
| 319 |
+
|
| 320 |
+
# -------------------------
|
| 321 |
+
# Application Setup
|
| 322 |
+
# -------------------------
|
| 323 |
+
|
| 324 |
+
model_manager = None
|
| 325 |
+
|
| 326 |
+
@asynccontextmanager
|
| 327 |
+
async def lifespan(app: FastAPI):
|
| 328 |
+
"""Application lifespan manager with model preloading."""
|
| 329 |
+
global model_manager
|
| 330 |
+
|
| 331 |
+
# Startup
|
| 332 |
+
logger.info("Starting reranking API...")
|
| 333 |
+
try:
|
| 334 |
+
model_manager = ModelManager()
|
| 335 |
+
await model_manager.preload_all_models()
|
| 336 |
+
logger.success("Reranking API startup complete!")
|
| 337 |
+
except Exception as e:
|
| 338 |
+
logger.error(f"Failed to initialize models: {e}")
|
| 339 |
+
raise
|
| 340 |
+
|
| 341 |
+
yield
|
| 342 |
+
|
| 343 |
+
# Shutdown
|
| 344 |
+
logger.info("Shutting down reranking API...")
|
| 345 |
+
|
| 346 |
+
|
| 347 |
+
app = FastAPI(
|
| 348 |
+
title="Multi-Model Reranking API",
|
| 349 |
+
description="""
|
| 350 |
+
High-performance API for document reranking using multiple state-of-the-art models.
|
| 351 |
+
|
| 352 |
+
β
**Supported Models:**
|
| 353 |
+
- **Jina Reranker V2**: Multilingual reranker optimized for search
|
| 354 |
+
- **BGE Reranker V2**: High-performance multilingual reranking
|
| 355 |
+
- **Qwen3 Reranker**: Instruction-based reranking with reasoning
|
| 356 |
+
|
| 357 |
+
π **Features:**
|
| 358 |
+
- Multiple reranking models preloaded at startup
|
| 359 |
+
- Batch document reranking with relevance scoring
|
| 360 |
+
- Optional instruction-based reranking (Qwen3)
|
| 361 |
+
- Comprehensive performance metrics
|
| 362 |
+
- Zero cold start delay
|
| 363 |
+
|
| 364 |
+
π **Input/Output:**
|
| 365 |
+
- Input: Query + documents + optional instruction
|
| 366 |
+
- Output: Ranked documents with relevance scores
|
| 367 |
+
""",
|
| 368 |
+
version="1.0.0",
|
| 369 |
+
lifespan=lifespan
|
| 370 |
+
)
|
| 371 |
+
|
| 372 |
+
|
| 373 |
+
# -------------------------
|
| 374 |
+
# API Endpoints
|
| 375 |
+
# -------------------------
|
| 376 |
+
|
| 377 |
+
@app.post("/rerank", response_model=RerankResponse, tags=["Reranking"])
|
| 378 |
+
async def rerank_documents(request: RerankRequest):
|
| 379 |
+
"""
|
| 380 |
+
Rerank documents based on relevance to query.
|
| 381 |
+
|
| 382 |
+
This endpoint takes a query and list of documents, then returns them
|
| 383 |
+
ranked by relevance using the specified reranking model.
|
| 384 |
+
|
| 385 |
+
Args:
|
| 386 |
+
request: RerankRequest containing query, documents, and model info
|
| 387 |
+
|
| 388 |
+
Returns:
|
| 389 |
+
RerankResponse with ranked documents, scores, and metadata
|
| 390 |
+
|
| 391 |
+
Example:
|
| 392 |
+
```json
|
| 393 |
+
{
|
| 394 |
+
"query": "machine learning algorithms",
|
| 395 |
+
"documents": [
|
| 396 |
+
"Deep learning uses neural networks",
|
| 397 |
+
"Weather forecast for tomorrow",
|
| 398 |
+
"Supervised learning with labeled data"
|
| 399 |
+
],
|
| 400 |
+
"model_id": "jina-reranker-v2"
|
| 401 |
+
}
|
| 402 |
+
```
|
| 403 |
+
"""
|
| 404 |
+
if not request.query.strip():
|
| 405 |
+
raise HTTPException(400, "Query cannot be empty")
|
| 406 |
+
|
| 407 |
+
if not request.documents:
|
| 408 |
+
raise HTTPException(400, "Documents list cannot be empty")
|
| 409 |
+
|
| 410 |
+
# Filter out empty documents
|
| 411 |
+
valid_docs = [(i, doc.strip()) for i, doc in enumerate(request.documents) if doc.strip()]
|
| 412 |
+
if not valid_docs:
|
| 413 |
+
raise HTTPException(400, "No valid documents found after filtering empty strings")
|
| 414 |
+
|
| 415 |
+
try:
|
| 416 |
+
start_time = time.time()
|
| 417 |
+
|
| 418 |
+
# Get model
|
| 419 |
+
model = model_manager.get_model(request.model_id)
|
| 420 |
+
|
| 421 |
+
# Extract valid documents and their indices
|
| 422 |
+
original_indices, documents = zip(*valid_docs)
|
| 423 |
+
|
| 424 |
+
# Perform reranking
|
| 425 |
+
scores = model.rerank(
|
| 426 |
+
query=request.query.strip(),
|
| 427 |
+
documents=list(documents),
|
| 428 |
+
instruction=request.instruction
|
| 429 |
+
)
|
| 430 |
+
|
| 431 |
+
# Create results with original indices
|
| 432 |
+
results = []
|
| 433 |
+
for i, (orig_idx, doc, score) in enumerate(zip(original_indices, documents, scores)):
|
| 434 |
+
results.append(RerankResult(
|
| 435 |
+
text=doc,
|
| 436 |
+
score=score,
|
| 437 |
+
index=orig_idx
|
| 438 |
+
))
|
| 439 |
+
|
| 440 |
+
# Sort by score (descending)
|
| 441 |
+
results.sort(key=lambda x: x.score, reverse=True)
|
| 442 |
+
|
| 443 |
+
# Apply top_k limit if specified
|
| 444 |
+
if request.top_k:
|
| 445 |
+
results = results[:request.top_k]
|
| 446 |
+
|
| 447 |
+
processing_time = time.time() - start_time
|
| 448 |
+
|
| 449 |
+
logger.info(
|
| 450 |
+
f"Reranked {len(documents)} documents in {processing_time:.3f}s "
|
| 451 |
+
f"using {request.model_id}"
|
| 452 |
+
)
|
| 453 |
+
|
| 454 |
+
return RerankResponse(
|
| 455 |
+
results=results,
|
| 456 |
+
query=request.query.strip(),
|
| 457 |
+
model_id=request.model_id,
|
| 458 |
+
processing_time=processing_time,
|
| 459 |
+
total_documents=len(request.documents),
|
| 460 |
+
returned_documents=len(results)
|
| 461 |
+
)
|
| 462 |
+
|
| 463 |
+
except ValueError as e:
|
| 464 |
+
raise HTTPException(400, str(e))
|
| 465 |
+
except Exception as e:
|
| 466 |
+
logger.error(f"Reranking failed: {e}")
|
| 467 |
+
raise HTTPException(500, f"Reranking failed: {str(e)}")
|
| 468 |
+
|
| 469 |
+
|
| 470 |
+
@app.get("/models", tags=["Models"])
|
| 471 |
+
async def list_models():
|
| 472 |
+
"""
|
| 473 |
+
List all available reranking models.
|
| 474 |
+
|
| 475 |
+
Returns information about all configured models including their
|
| 476 |
+
loading status and capabilities.
|
| 477 |
+
|
| 478 |
+
Returns:
|
| 479 |
+
List of model information dictionaries
|
| 480 |
+
"""
|
| 481 |
+
try:
|
| 482 |
+
return model_manager.list_models()
|
| 483 |
+
except Exception as e:
|
| 484 |
+
logger.error(f"Failed to list models: {e}")
|
| 485 |
+
raise HTTPException(500, str(e))
|
| 486 |
+
|
| 487 |
+
|
| 488 |
+
@app.get("/health", tags=["Monitoring"])
|
| 489 |
+
async def health_check():
|
| 490 |
+
"""
|
| 491 |
+
Check API health and model status.
|
| 492 |
+
|
| 493 |
+
Returns comprehensive health information including model loading
|
| 494 |
+
status and system metrics.
|
| 495 |
+
|
| 496 |
+
Returns:
|
| 497 |
+
Health status dictionary
|
| 498 |
+
"""
|
| 499 |
+
try:
|
| 500 |
+
models = model_manager.list_models()
|
| 501 |
+
loaded_models = [m for m in models if m['loaded']]
|
| 502 |
+
|
| 503 |
+
return {
|
| 504 |
+
"status": "ok",
|
| 505 |
+
"total_models": len(models),
|
| 506 |
+
"loaded_models": len(loaded_models),
|
| 507 |
+
"available_models": [m['id'] for m in loaded_models],
|
| 508 |
+
"models_info": models
|
| 509 |
+
}
|
| 510 |
+
except Exception as e:
|
| 511 |
+
logger.error(f"Health check failed: {e}")
|
| 512 |
+
return {
|
| 513 |
+
"status": "error",
|
| 514 |
+
"error": str(e)
|
| 515 |
+
}
|
| 516 |
|
|
|
|
|
|
|
|
|
requirements.txt
CHANGED
|
@@ -1,2 +1,6 @@
|
|
| 1 |
-
fastapi
|
| 2 |
-
uvicorn[standard]
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
fastapi==0.116.2
|
| 2 |
+
uvicorn[standard]==0.35.0
|
| 3 |
+
torch==2.8.0
|
| 4 |
+
sentence-transformers==5.1.1
|
| 5 |
+
loguru==0.7.3
|
| 6 |
+
einops==0.8.1
|