Spaces:
Running
Running
File size: 8,076 Bytes
d557d77 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 |
"""Response generation using local or remote LLMs."""
from typing import Optional
from coderag.config import get_settings
from coderag.generation.citations import CitationParser
from coderag.generation.prompts import SYSTEM_PROMPT, build_prompt, build_no_context_response
from coderag.logging import get_logger
from coderag.models.response import Response
from coderag.models.query import Query
from coderag.retrieval.retriever import Retriever
logger = get_logger(__name__)
class ResponseGenerator:
"""Generates grounded responses using local or remote LLMs."""
def __init__(
self,
retriever: Optional[Retriever] = None,
) -> None:
self.settings = get_settings()
self.retriever = retriever or Retriever()
self.citation_parser = CitationParser()
self.provider = self.settings.models.llm_provider.lower()
self._client = None
self._local_model = None
self._local_tokenizer = None
logger.info("ResponseGenerator initialized", provider=self.provider)
def _get_api_client(self):
"""Get or create API client for remote providers."""
if self._client is not None:
return self._client
import httpx
from openai import OpenAI
api_key = self.settings.models.llm_api_key
if not api_key:
raise ValueError(f"API key required for provider: {self.provider}")
# Provider-specific configurations
provider_configs = {
"openai": {
"base_url": "https://api.openai.com/v1",
"default_model": "gpt-4o-mini",
},
"groq": {
"base_url": "https://api.groq.com/openai/v1",
"default_model": "llama-3.3-70b-versatile",
},
"anthropic": {
"base_url": "https://api.anthropic.com/v1",
"default_model": "claude-3-5-sonnet-20241022",
},
"openrouter": {
"base_url": "https://openrouter.ai/api/v1",
"default_model": "anthropic/claude-3.5-sonnet",
},
"together": {
"base_url": "https://api.together.xyz/v1",
"default_model": "meta-llama/Llama-3.3-70B-Instruct-Turbo",
},
}
config = provider_configs.get(self.provider, {})
base_url = self.settings.models.llm_api_base or config.get("base_url")
if not base_url:
raise ValueError(f"Unknown provider: {self.provider}")
# Set default model if not specified and it's a known provider
if self.settings.models.llm_name.startswith("Qwen/"):
self.model_name = config.get("default_model", self.settings.models.llm_name)
else:
self.model_name = self.settings.models.llm_name
self._client = OpenAI(
api_key=api_key,
base_url=base_url,
http_client=httpx.Client(timeout=120.0),
)
logger.info("API client created", provider=self.provider, model=self.model_name)
return self._client
def _load_local_model(self):
"""Load local model with transformers."""
if self._local_model is not None:
return
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig
if not torch.cuda.is_available():
raise RuntimeError(
"Local LLM requires a CUDA-capable GPU. Options:\n"
" 1. Use a cloud provider (free): MODEL_LLM_PROVIDER=groq\n"
" Get API key at: https://console.groq.com/keys\n"
" 2. Install CUDA and a compatible GPU"
)
logger.info("Loading local LLM", model=self.settings.models.llm_name)
if self.settings.models.llm_use_4bit:
bnb_config = BitsAndBytesConfig(
load_in_4bit=True,
bnb_4bit_quant_type="nf4",
bnb_4bit_compute_dtype=torch.float16,
bnb_4bit_use_double_quant=True,
)
else:
bnb_config = None
self._local_tokenizer = AutoTokenizer.from_pretrained(
self.settings.models.llm_name,
trust_remote_code=True,
)
self._local_model = AutoModelForCausalLM.from_pretrained(
self.settings.models.llm_name,
quantization_config=bnb_config,
device_map=self.settings.models.llm_device_map,
trust_remote_code=True,
torch_dtype=torch.float16,
)
logger.info("Local LLM loaded successfully")
def generate(self, query: Query) -> Response:
"""Generate a response for a query."""
# Retrieve relevant chunks
chunks, context = self.retriever.retrieve_with_context(
query.question,
query.repo_id,
query.top_k,
)
# Handle no results
if not chunks:
return Response(
answer=build_no_context_response(),
citations=[],
retrieved_chunks=[],
grounded=False,
query_id=query.id,
)
# Build prompt and generate
prompt = build_prompt(query.question, context)
if self.provider == "local":
answer = self._generate_local(prompt)
else:
answer = self._generate_api(prompt)
# Parse citations from answer
citations = self.citation_parser.parse_citations(answer)
# Determine if response is grounded
grounded = len(citations) > 0 and len(chunks) > 0
return Response(
answer=answer,
citations=citations,
retrieved_chunks=chunks,
grounded=grounded,
query_id=query.id,
)
def _generate_api(self, prompt: str) -> str:
"""Generate using remote API."""
client = self._get_api_client()
messages = [
{"role": "system", "content": SYSTEM_PROMPT},
{"role": "user", "content": prompt},
]
response = client.chat.completions.create(
model=self.model_name,
messages=messages,
max_tokens=self.settings.models.llm_max_new_tokens,
temperature=self.settings.models.llm_temperature,
top_p=self.settings.models.llm_top_p,
)
return response.choices[0].message.content.strip()
def _generate_local(self, prompt: str) -> str:
"""Generate using local model."""
import torch
self._load_local_model()
messages = [
{"role": "system", "content": SYSTEM_PROMPT},
{"role": "user", "content": prompt},
]
text = self._local_tokenizer.apply_chat_template(
messages,
tokenize=False,
add_generation_prompt=True,
)
inputs = self._local_tokenizer(text, return_tensors="pt").to(self._local_model.device)
with torch.no_grad():
outputs = self._local_model.generate(
**inputs,
max_new_tokens=self.settings.models.llm_max_new_tokens,
temperature=self.settings.models.llm_temperature,
top_p=self.settings.models.llm_top_p,
do_sample=True,
pad_token_id=self._local_tokenizer.eos_token_id,
)
generated = outputs[0][inputs["input_ids"].shape[1]:]
response = self._local_tokenizer.decode(generated, skip_special_tokens=True)
return response.strip()
def unload(self) -> None:
"""Unload models from memory."""
if self._local_model is not None:
del self._local_model
self._local_model = None
if self._local_tokenizer is not None:
del self._local_tokenizer
self._local_tokenizer = None
import torch
if torch.cuda.is_available():
torch.cuda.empty_cache()
logger.info("Models unloaded")
|