File size: 15,361 Bytes
e0b68d2 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 | #!/usr/bin/env python
# -*- coding: utf-8 -*-
"""Minimal SSE response handler.
Converts upstream SSE responses to OpenAI-compatible format.
"""
import json
import time
import uuid
from dataclasses import dataclass, field
from typing import Any, AsyncGenerator, Dict, List, Optional, Tuple
import httpx
from app.core.openai_compat import (
create_openai_chunk,
create_openai_response_with_reasoning,
format_sse_chunk,
format_sse_done,
get_error_message,
handle_error,
)
from app.utils.logger import logger, log_exception
# ------------------------------------------------------------------
# UpstreamChunkEvent
# ------------------------------------------------------------------
@dataclass
class UpstreamChunkEvent:
delta_content: str = ""
edit_content: str = ""
edit_index: Optional[int] = None
phase: str = ""
done: bool = False
usage: Optional[Dict] = None
error: Optional[Dict] = None
data: Dict = field(default_factory=dict)
chunk_type: str = ""
raw_chunk: Optional[Dict] = None
def _parse_chunk_event(chunk: dict) -> Optional[UpstreamChunkEvent]:
if not isinstance(chunk, dict):
return None
chunk_type = str(chunk.get("type", ""))
data: Any = chunk.get("data", {}) if chunk_type == "chat:completion" else chunk
if not isinstance(data, dict):
return None
error = None
if chunk.get("error"):
err = chunk["error"]
error = err if isinstance(err, dict) else {"detail": str(err), "code": "UNKNOWN"}
elif data.get("error"):
err = data["error"]
error = err if isinstance(err, dict) else {"detail": str(err), "code": "UNKNOWN"}
elif isinstance(data.get("inner"), dict) and data["inner"].get("error"):
err = data["inner"]["error"]
error = err if isinstance(err, dict) else {"detail": str(err), "code": "UNKNOWN"}
return UpstreamChunkEvent(
delta_content=str(data.get("delta_content", "") or ""),
edit_content=str(data.get("edit_content", "") or ""),
edit_index=data.get("edit_index") if isinstance(data.get("edit_index"), int) else None,
phase=str(data.get("phase", "") or ""),
done=bool(data.get("done") is True),
usage=(
data.get("usage") if isinstance(data.get("usage"), dict)
else chunk.get("usage") if isinstance(chunk.get("usage"), dict)
else None
),
error=error,
data=data,
chunk_type=chunk_type,
raw_chunk=chunk,
)
# ------------------------------------------------------------------
# StreamContext
# ------------------------------------------------------------------
@dataclass
class StreamContext:
chat_id: str
model: str
stream_id: Optional[str] = None
created_at: Optional[int] = None
buffered_content: str = ""
usage_info: Dict[str, Any] = field(
default_factory=lambda: {"prompt_tokens": 0, "completion_tokens": 0, "total_tokens": 0}
)
has_sent_role: bool = False
finished: bool = False
line_count: int = 0
last_phase: Optional[str] = None
def ensure_stream_id(self, chunk_data: Optional[Dict] = None) -> str:
if self.stream_id is None:
upstream_id = chunk_data.get("id") if isinstance(chunk_data, dict) else None
self.stream_id = upstream_id or f"chatcmpl-{uuid.uuid4().hex[:12]}"
self.created_at = int(time.time())
return self.stream_id
# ------------------------------------------------------------------
# ResponseHandler
# ------------------------------------------------------------------
class ResponseHandler:
# SSE helpers
def _emit_sse(self, ctx: StreamContext, delta: Dict[str, Any]) -> List[str]:
output: List[str] = []
role_sse = self._ensure_role_sse(ctx)
if role_sse:
output.append(role_sse)
output.append(format_sse_chunk(
create_openai_chunk(ctx.ensure_stream_id(), ctx.model, delta, created=ctx.created_at)
))
return output
def _terminal_sse(self, ctx: StreamContext, *, finish_reason: str) -> List[str]:
chunk = create_openai_chunk(ctx.ensure_stream_id(), ctx.model, {}, finish_reason, created=ctx.created_at)
chunk["usage"] = ctx.usage_info
return [format_sse_chunk(chunk), format_sse_done()]
def _ensure_role_sse(self, ctx: StreamContext) -> Optional[str]:
if ctx.has_sent_role:
return None
ctx.has_sent_role = True
return format_sse_chunk(
create_openai_chunk(ctx.ensure_stream_id(), ctx.model, {"role": "assistant"}, created=ctx.created_at)
)
# SSE line parsing
@staticmethod
def _extract_sse_payload(line: str) -> Tuple[str, str]:
if not line:
return "skip", ""
if line.startswith("data:"):
payload = line[5:].strip()
else:
stripped = line.strip()
if not stripped:
return "skip", ""
if not stripped.startswith("data:"):
return "non_data", stripped
payload = stripped[5:].strip()
if not payload:
return "skip", ""
if payload == "[DONE]" or payload.casefold() == "done":
return "done", ""
return "data", payload
def _parse_sse_line(self, line: str, ctx: StreamContext) -> Tuple[str, Optional[UpstreamChunkEvent]]:
if not line.strip():
return "skip", None
ctx.line_count += 1
state, chunk_str = self._extract_sse_payload(line)
if state == "done":
return "done", None
if state != "data":
return "skip", None
try:
chunk = json.loads(chunk_str)
except json.JSONDecodeError:
return "skip", None
if not isinstance(chunk, dict):
return "skip", None
event = _parse_chunk_event(chunk)
return ("ok", event) if event else ("skip", None)
# State update
def _update_state(self, ctx: StreamContext, event: UpstreamChunkEvent) -> None:
ctx.ensure_stream_id(event.raw_chunk)
phase = event.phase
if phase and phase != ctx.last_phase:
ctx.last_phase = phase
if event.usage and event.usage != ctx.usage_info:
ctx.usage_info = event.usage
# Content accumulation (needed for edit_index insertion)
def _accumulate(self, ctx: StreamContext, event: UpstreamChunkEvent) -> str:
max_len = 50000
if event.delta_content:
ctx.buffered_content = (ctx.buffered_content + event.delta_content)[-max_len:]
return event.delta_content
if event.edit_content:
idx = event.edit_index
if idx is not None and isinstance(idx, int):
safe = max(0, min(idx, len(ctx.buffered_content)))
ctx.buffered_content = ctx.buffered_content[:safe] + event.edit_content + ctx.buffered_content[safe:]
if len(ctx.buffered_content) > max_len:
ctx.buffered_content = ctx.buffered_content[-max_len:]
else:
ctx.buffered_content = (ctx.buffered_content + event.edit_content)[-max_len:]
return event.edit_content
return ""
# Error check
def _check_error(self, ctx: StreamContext, event: UpstreamChunkEvent) -> Optional[List[str]]:
if not event.error:
return None
code = event.error.get("code", "UNKNOWN")
detail = get_error_message(Exception(str(event.error.get("detail", "Unknown upstream error"))))
if event.done:
return None
logger.error("upstream SSE error: code=%s, detail=%s", code, detail)
output: List[str] = []
role = self._ensure_role_sse(ctx)
if role:
output.append(role)
output.append(f'data: {json.dumps({"error": {"message": detail, "type": "upstream_error", "code": code}})}\n\n')
output.append("data: [DONE]\n\n")
return output
# Search results
def _format_search(self, data: Dict[str, Any]) -> str:
info = data.get("results") or data.get("sources") or data.get("citations")
if not isinstance(info, list) or not info:
return ""
parts = []
for i, item in enumerate(info, 1):
if not isinstance(item, dict):
continue
title = item.get("title") or item.get("name") or f"Result {i}"
url = item.get("url") or item.get("link")
if url:
parts.append(f"[{i}] [{title}]({url})")
return ("\n\n---\n" + "\n".join(parts)) if parts else ""
# Phase output
def _emit_phase(self, ctx: StreamContext, event: UpstreamChunkEvent, text: str) -> List[str]:
phase = ctx.last_phase
if phase == "thinking" and event.delta_content:
return self._emit_sse(ctx, {"reasoning_content": event.delta_content})
if phase in ("answer", "other") and text:
return self._emit_sse(ctx, {"content": text})
if phase == "search" or event.chunk_type == "web_search":
citation = self._format_search(event.data)
if citation:
return self._emit_sse(ctx, {"content": citation})
return []
# Non-stream edit helper
@staticmethod
def _apply_edit(parts: List[str], edit: str, idx: Optional[int]) -> None:
if not edit:
return
if idx is not None and isinstance(idx, int):
joined = "".join(parts)
safe = max(0, min(idx, len(joined)))
parts[:] = [joined[:safe] + edit + joined[safe:]]
else:
parts.append(edit)
# ==================================================================
# Stream response (main entry)
# ==================================================================
async def handle_stream_response(
self,
response: httpx.Response,
chat_id: str,
model: str,
started_at: Optional[float] = None,
) -> AsyncGenerator[str, None]:
if started_at:
logger.debug("[stream] start SSE, ttfb: %.3fs", time.perf_counter() - started_at)
sse_start = time.perf_counter()
ctx = StreamContext(chat_id=chat_id, model=model)
try:
async for line in response.aiter_lines():
state, event = self._parse_sse_line(line, ctx)
if state == "done":
break
if state != "ok" or event is None:
continue
err = self._check_error(ctx, event)
if err:
for sse in err:
yield sse
return
self._update_state(ctx, event)
text = self._accumulate(ctx, event)
for sse in self._emit_phase(ctx, event, text):
yield sse
if event.done:
break
if not ctx.finished:
role = self._ensure_role_sse(ctx)
if role:
yield role
for sse in self._terminal_sse(ctx, finish_reason="stop"):
yield sse
ctx.finished = True
except Exception as e:
log_exception(logger, "stream error")
if not ctx.finished:
yield format_sse_chunk(
create_openai_chunk(ctx.ensure_stream_id(), model, {}, "stop", created=ctx.created_at)
)
yield format_sse_done()
finally:
elapsed = time.perf_counter() - sse_start
total = f", total: {time.perf_counter() - started_at:.3f}s" if started_at else ""
logger.debug("[stream] done %s lines, %.3fs%s", ctx.line_count, elapsed, total)
# ==================================================================
# Non-stream response
# ==================================================================
async def handle_non_stream_response(
self,
response: httpx.Response,
chat_id: str,
model: str,
) -> Dict[str, Any]:
content_parts: List[str] = []
reasoning_parts: List[str] = []
usage: Dict[str, Any] = {"prompt_tokens": 0, "completion_tokens": 0, "total_tokens": 0}
last_phase = None
try:
async for line in response.aiter_lines():
state, payload = self._extract_sse_payload(line)
if state == "non_data":
try:
err = json.loads(payload)
if isinstance(err, dict) and ("error" in err or "code" in err or "message" in err):
msg = (
(err.get("error") or {}).get("message")
if isinstance(err.get("error"), dict)
else err.get("message")
) or "upstream error"
return handle_error(Exception(msg), "API response")
except Exception:
pass
continue
if state != "data":
continue
try:
chunk = json.loads(payload)
except json.JSONDecodeError:
continue
event = _parse_chunk_event(chunk)
if event is None:
continue
phase = event.phase
if phase and phase != last_phase:
last_phase = phase
if phase == "thinking" and event.delta_content:
reasoning_parts.append(event.delta_content)
elif phase == "answer":
if event.delta_content:
content_parts.append(event.delta_content)
elif event.edit_content:
self._apply_edit(content_parts, event.edit_content, event.edit_index)
elif phase == "other" and event.edit_content:
self._apply_edit(content_parts, event.edit_content, event.edit_index)
elif phase == "search" or event.chunk_type == "web_search":
s = self._format_search(event.data)
if s:
content_parts.append(s)
if event.usage:
usage = event.usage
except Exception as e:
log_exception(logger, "non-stream error")
return handle_error(e, "non-stream aggregation")
final = "".join(content_parts).strip()
reasoning = "".join(reasoning_parts).strip()
if not final and reasoning:
final = reasoning
return create_openai_response_with_reasoning(chat_id, model, final, reasoning, usage)
|