"""Expose one Query turn as server-sent events with same-turn error recovery.""" import asyncio import contextvars import inspect import json import threading import zlib import anyio from fastapi import APIRouter, Body, Depends from starlette.responses import StreamingResponse from deps import Session, module_gate, require_session import provider_limits import provider_stream PATH = "/api/v1/query/chat/stream" router = APIRouter(prefix="/api/v1", dependencies=[Depends(module_gate("analyst"))]) class StreamGZipMiddleware: """Compress response chunks in the bounded anyio pool while passing SSE through.""" def __init__(self, app, **kwargs): self.app = app self.minimum_size = kwargs.get("minimum_size", 500) self.compresslevel = kwargs.get("compresslevel", 9) async def __call__(self, scope, receive, send): if scope.get("type") != "http" or scope.get("path") == PATH: return await self.app(scope, receive, send) from starlette.datastructures import Headers from starlette.middleware.gzip import IdentityResponder if "gzip" in Headers(scope=scope).get("Accept-Encoding", ""): responder = _OffLoopGZipResponder(self.app, self.minimum_size, self.compresslevel) else: responder = IdentityResponder(self.app, self.minimum_size) await responder(scope, receive, send) from starlette.middleware.gzip import IdentityResponder class _OffLoopGZipResponder(IdentityResponder): """Retain Starlette's header and chunk behavior with compression outside the event loop.""" content_encoding = "gzip" def __init__(self, app, minimum_size, compresslevel): super().__init__(app, minimum_size) self.compresslevel = compresslevel self._compressor = None def _compress(self, body, more): if self._compressor is None: self._compressor = zlib.compressobj(self.compresslevel, zlib.DEFLATED, 16 + zlib.MAX_WBITS) out = self._compressor.compress(body) return out if more else out + self._compressor.flush() async def send_with_compression(self, message): if (message["type"] == "http.response.body" and not self.content_encoding_set and not self.content_type_is_excluded): body, more = message.get("body", b""), message.get("more_body", False) if self.started or more or len(body) >= self.minimum_size: self._compressed_body = await anyio.to_thread.run_sync( lambda: self._compress(body, more)) await super().send_with_compression(message) # Starlette 1.4 awaits apply_compression and 1.3 calls it, so match the installed base. if inspect.iscoroutinefunction(IdentityResponder.apply_compression): async def apply_compression(self, body, *, more_body): return self._compressed_body else: def apply_compression(self, body, *, more_body): return self._compressed_body def _event(name, data): return "event: " + name + "\ndata: " + json.dumps(data, ensure_ascii=False) + "\n\n" async def events(body, session, *, submit=None): """Run _submit once and send its result or same-turn recovery through the event queue.""" import routes_query submit = submit or routes_query._submit loop = asyncio.get_running_loop() queue = asyncio.Queue() emitted = [] failed = [] prefix_sent = False connected = True def put(name, data): if connected: loop.call_soon_threadsafe(queue.put_nowait, (name, data)) def hook(name, data): nonlocal prefix_sent if name == "stream_failure": failed.append(data["code"]) return if name == "thinking": handle = (routes_query._AGENT_TURN.get() or {}).get("handle") if handle is not None: data = {**data, "turnId": handle.id} if name == "text_delta": turn = routes_query._AGENT_TURN.get() or {} handle = turn.get("handle") if handle is not None and (handle.stopped or handle.cancel_requested.is_set()): return if turn.get("providerPaused") and not prefix_sent and not emitted: prefix = provider_limits.RESUMED_MESSAGE + "\n\n" emitted.append(prefix) put("text_delta", {"text": prefix}) prefix_sent = True emitted.append(data["text"]) put(name, data) def work(): token = provider_stream.EVENT_HOOK.set(hook) try: result = submit(body, session) if result.get("status") == "joined": put("joined", result) elif result.get("status") == "stopped": put("error", {"code": "turn_stopped", "result": result, "retry": False}) else: content = result["message"]["content"] text = "".join(emitted) if failed or (text and content != text): put("error", {"code": failed[0] if failed else "stream_replaced", "result": result, "retry": False}) else: if content[len(text):]: put("text_delta", {"text": content[len(text):]}) put("done", result) except Exception as exc: from request_timing import resource_cause cause = resource_cause(exc) if cause is not None: put("error", {"code": "server_busy", "cause": cause, "message": "The server is busy. Try again shortly.", "retry": False}) return detail = getattr(exc, "detail", None) error = (detail or {}).get("error") if isinstance(detail, dict) else None put("error", {"code": (error or {}).get("code", "stream_failed"), "message": (error or {}).get("message", "The turn could not finish."), "retry": False}) finally: provider_stream.EVENT_HOOK.reset(token) put("_end", None) context = contextvars.copy_context() worker = threading.Thread(target=context.run, args=(work,), daemon=True, name="assistant-stream") worker.start() try: while True: name, data = await queue.get() if name == "_end": break yield _event(name, data) finally: # Disconnect does not replay or cancel writes in an already admitted turn. connected = False @router.post("/query/chat/stream") def stream_chat(body: dict = Body(default=None), session: Session = Depends(require_session)): return StreamingResponse(events(body, session), media_type="text/event-stream", headers={"Cache-Control": "no-cache", "X-Accel-Buffering": "no"})