langgraph-api 0.0.1__py3-none-any.whl

langgraph_api/js/remote.py

@@ -0,0 +1,673 @@
+import asyncio
+import os
+import shutil
+from collections.abc import AsyncIterator
+from typing import Any, Literal
+
+import httpx
+import orjson
+import structlog
+import uvicorn
+from langchain_core.runnables import Runnable
+from langchain_core.runnables.config import RunnableConfig
+from langchain_core.runnables.graph import Edge, Node
+from langchain_core.runnables.graph import Graph as DrawableGraph
+from langchain_core.runnables.schema import (
+    CustomStreamEvent,
+    StandardStreamEvent,
+    StreamEvent,
+)
+from langgraph.checkpoint.serde.base import SerializerProtocol
+from langgraph.pregel.types import PregelTask, StateSnapshot
+from langgraph.store.base import GetOp, Item, ListNamespacesOp, PutOp, SearchOp
+from langgraph.types import Interrupt
+from pydantic import BaseModel
+from starlette.applications import Starlette
+from starlette.requests import Request
+from starlette.routing import Route
+
+from langgraph_api.js.server_sent_events import aconnect_sse
+from langgraph_api.route import ApiResponse
+from langgraph_api.serde import json_dumpb
+from langgraph_api.utils import AsyncConnectionProto
+
+logger = structlog.stdlib.get_logger(__name__)
+
+GRAPH_SOCKET = "./graph.sock"
+CHECKPOINTER_SOCKET = "./checkpointer.sock"
+STORE_SOCKET = "./store.sock"
+
+
+class NoopModel(BaseModel):
+    pass
+
+
+class RemoteException(Exception):
+    error: str
+
+    def __init__(self, error: str, *args: object) -> None:
+        super().__init__(*args)
+        self.error = error
+
+    # Used to nudge the serde to encode like BaseException
+    # @see /api/langgraph_api/shared/serde.py:default
+    def dict(self):
+        return {"error": self.error, "message": str(self)}
+
+
+# Shim for the Pregel API. Will connect to GRAPH_SOCKET
+# UNIX socket to communicate with the JS process.
+class RemotePregel(Runnable):
+    # TODO: implement name overriding
+    name: str = "LangGraph"
+
+    # TODO: implement graph_id overriding
+    graph_id: str
+
+    _async_client: httpx.AsyncClient
+
+    @staticmethod
+    async def load(graph_id: str):
+        model = RemotePregel()
+
+        model.graph_id = graph_id
+        model._async_client = httpx.AsyncClient(
+            base_url="http://graph",
+            timeout=httpx.Timeout(None),
+            limits=httpx.Limits(),
+            transport=httpx.AsyncHTTPTransport(uds=GRAPH_SOCKET),
+        )
+
+        return model
+
+    async def astream_events(
+        self,
+        input: Any,
+        config: RunnableConfig | None = None,
+        *,
+        version: Literal["v1", "v2"],
+        **kwargs: Any,
+    ) -> AsyncIterator[StreamEvent]:
+        if version != "v2":
+            raise ValueError("Only v2 of astream_events is supported")
+
+        async with aconnect_sse(
+            self._async_client,
+            "POST",
+            f"/{self.graph_id}/streamEvents",
+            headers={"Content-Type": "application/json"},
+            data=orjson.dumps({"input": input, "config": config, **kwargs}),
+        ) as event_source:
+            async for sse in event_source.aiter_sse():
+                event = orjson.loads(sse["data"])
+                if sse["event"] == "error":
+                    raise RemoteException(event["error"], event["message"])
+                elif event["event"] == "on_custom_event":
+                    yield CustomStreamEvent(**event)
+                else:
+                    yield StandardStreamEvent(**event)
+
+    async def fetch_state_schema(self):
+        schema = await self._async_client.post(f"/{self.graph_id}/getSchema")
+        return orjson.loads(schema.content)
+
+    async def fetch_graph(
+        self,
+        config: RunnableConfig | None = None,
+        *,
+        xray: int | bool = False,
+    ) -> DrawableGraph:
+        response = (
+            await self._async_client.post(
+                f"/{self.graph_id}/getGraph",
+                headers={"Content-Type": "application/json"},
+                data=orjson.dumps({"config": config, "xray": xray}),
+            )
+        ).json()
+
+        nodes: list[Any] = response.pop("nodes")
+        edges: list[Any] = response.pop("edges")
+
+        return DrawableGraph(
+            {
+                data["id"]: Node(
+                    data["id"], data["id"], NoopModel(), data.get("metadata")
+                )
+                for data in nodes
+            },
+            {
+                Edge(
+                    data["source"],
+                    data["target"],
+                    data.get("data"),
+                    data.get("conditional", False),
+                )
+                for data in edges
+            },
+        )
+
+    async def fetch_subgraphs(
+        self, *, namespace: str | None = None, recurse: bool = False
+    ) -> dict[str, dict]:
+        return (
+            await self._async_client.post(
+                f"/{self.graph_id}/getSubgraphs",
+                headers={"Content-Type": "application/json"},
+                data=orjson.dumps({"namespace": namespace, "recurse": recurse}),
+            )
+        ).json()
+
+    def _convert_state_snapshot(self, item: dict) -> StateSnapshot:
+        def _convert_tasks(tasks: list[dict]) -> tuple[PregelTask, ...]:
+            result: list[PregelTask] = []
+            for task in tasks:
+                state = task.get("state")
+
+                if state and isinstance(state, dict) and "config" in state:
+                    state = self._convert_state_snapshot(state)
+
+                result.append(
+                    PregelTask(
+                        task["id"],
+                        task["name"],
+                        tuple(task["path"]) if task.get("path") else tuple(),
+                        # TODO: figure out how to properly deserialise errors
+                        task.get("error"),
+                        tuple(
+                            Interrupt(
+                                value=interrupt["value"],
+                                when=interrupt["when"],
+                            )
+                            for interrupt in task.get("interrupts")
+                        )
+                        if task.get("interrupts")
+                        else [],
+                        state,
+                    )
+                )
+            return tuple(result)
+
+        return StateSnapshot(
+            item.get("values"),
+            item.get("next"),
+            item.get("config"),
+            item.get("metadata"),
+            item.get("createdAt"),
+            item.get("parentConfig"),
+            _convert_tasks(item.get("tasks", [])),
+        )
+
+    async def aget_state(
+        self, config: RunnableConfig, *, subgraphs: bool = False
+    ) -> StateSnapshot:
+        response = await self._async_client.post(
+            f"/{self.graph_id}/getState",
+            headers={"Content-Type": "application/json"},
+            data=orjson.dumps({"config": config, "subgraphs": subgraphs}),
+        )
+        return self._convert_state_snapshot(response.json())
+
+    async def aupdate_state(
+        self,
+        config: RunnableConfig,
+        values: dict[str, Any] | Any,
+        as_node: str | None = None,
+    ) -> RunnableConfig:
+        response = await self._async_client.post(
+            f"/{self.graph_id}/updateState",
+            headers={"Content-Type": "application/json"},
+            data=orjson.dumps({"config": config, "values": values, "as_node": as_node}),
+        )
+        return RunnableConfig(**response.json())
+
+    async def aget_state_history(
+        self,
+        config: RunnableConfig,
+        *,
+        filter: dict[str, Any] | None = None,
+        before: RunnableConfig | None = None,
+        limit: int | None = None,
+    ) -> AsyncIterator[StateSnapshot]:
+        async with aconnect_sse(
+            self._async_client,
+            "POST",
+            f"/{self.graph_id}/getStateHistory",
+            headers={"Content-Type": "application/json"},
+            data=orjson.dumps(
+                {"config": config, "limit": limit, "filter": filter, "before": before}
+            ),
+        ) as event_source:
+            async for sse in event_source.aiter_sse():
+                yield self._convert_state_snapshot(orjson.loads(sse["data"]))
+
+    def get_graph(
+        self,
+        config: RunnableConfig | None = None,
+        *,
+        xray: int | bool = False,
+    ) -> dict[str, Any]:
+        raise Exception("Not implemented")
+
+    def get_input_schema(self, config: RunnableConfig | None = None) -> type[BaseModel]:
+        raise Exception("Not implemented")
+
+    def get_output_schema(
+        self, config: RunnableConfig | None = None
+    ) -> type[BaseModel]:
+        raise Exception("Not implemented")
+
+    def config_schema(self) -> type[BaseModel]:
+        raise Exception("Not implemented")
+
+    async def invoke(self, input: Any, config: RunnableConfig | None = None):
+        raise Exception("Not implemented")
+
+
+async def run_js_process(paths_str: str, watch: bool = False):
+    # check if tsx is available
+    tsx_path = shutil.which("tsx")
+    if tsx_path is None:
+        raise FileNotFoundError("tsx not found in PATH")
+    attempt = 0
+    while True:
+        client_file = os.path.join(os.path.dirname(__file__), "client.mts")
+        args = ("tsx", client_file)
+        if watch:
+            args = ("tsx", "watch", client_file, "--skip-schema-cache")
+        try:
+            process = await asyncio.create_subprocess_exec(
+                *args,
+                env={
+                    "LANGSERVE_GRAPHS": paths_str,
+                    "LANGCHAIN_CALLBACKS_BACKGROUND": "true",
+                    "CHOKIDAR_USEPOLLING": "true",
+                    **os.environ,
+                },
+            )
+            code = await process.wait()
+            raise Exception(f"JS process exited with code {code}")
+        except asyncio.CancelledError:
+            logger.info("Terminating JS graphs process")
+            try:
+                process.terminate()
+                await process.wait()
+            except (UnboundLocalError, ProcessLookupError):
+                pass
+            raise
+        except Exception:
+            if attempt >= 3:
+                raise
+            else:
+                logger.warning(f"Retrying JS process {3 - attempt} more times...")
+                attempt += 1
+
+
+def _get_passthrough_checkpointer(conn: AsyncConnectionProto):
+    from langgraph_storage.checkpoint import Checkpointer
+
+    class PassthroughSerialiser(SerializerProtocol):
+        def dumps(self, obj: Any) -> bytes:
+            return json_dumpb(obj)
+
+        def dumps_typed(self, obj: Any) -> tuple[str, bytes]:
+            return "json", json_dumpb(obj)
+
+        def loads(self, data: bytes) -> Any:
+            return orjson.loads(data)
+
+        def loads_typed(self, data: tuple[str, bytes]) -> Any:
+            type, payload = data
+            if type != "json":
+                raise ValueError(f"Unsupported type {type}")
+            return orjson.loads(payload)
+
+    checkpointer = Checkpointer(conn)
+
+    # This checkpointer does not attempt to revive LC-objects.
+    # Instead, it will pass through the JSON values as-is.
+    checkpointer.serde = PassthroughSerialiser()
+
+    return checkpointer
+
+
+# Setup a HTTP server on top of CHECKPOINTER_SOCKET unix socket
+# used by `client.mts` to communicate with the Python checkpointer
+async def run_remote_checkpointer():
+    from langgraph_storage.database import connect
+
+    # Search checkpoints
+    async def list(request: Request):
+        payload = orjson.loads(await request.body())
+        result = []
+        async with connect() as conn:
+            checkpointer = _get_passthrough_checkpointer(conn)
+            async for item in checkpointer.alist(
+                config=payload.get("config"),
+                limit=payload.get("limit"),
+                before=payload.get("before"),
+                filter=payload.get("filter"),
+            ):
+                result.append(item)
+
+        return ApiResponse(result)
+
+    # Put the new checkpoint metadata
+    async def put(request: Request):
+        payload = orjson.loads(await request.body())
+        async with connect() as conn:
+            checkpointer = _get_passthrough_checkpointer(conn)
+            result = await checkpointer.aput(
+                payload["config"],
+                payload["checkpoint"],
+                payload["metadata"],
+                payload.get("new_versions", {}),
+            )
+        return ApiResponse(result)
+
+    # Get actual checkpoint values (reads)
+    async def get_tuple(request: Request):
+        payload = orjson.loads(await request.body())
+
+        async with connect() as conn:
+            checkpointer = _get_passthrough_checkpointer(conn)
+            result = await checkpointer.aget_tuple(config=payload["config"])
+        return ApiResponse(result)
+
+    # Put actual checkpoint values (writes)
+    async def put_writes(request: Request):
+        payload = orjson.loads(await request.body())
+
+        async with connect() as conn:
+            checkpointer = _get_passthrough_checkpointer(conn)
+            result = await checkpointer.aput_writes(
+                payload["config"],
+                payload["writes"],
+                payload["taskId"],
+            )
+
+        return ApiResponse(result)
+
+    remote = Starlette(
+        routes=[
+            Route("/get_tuple", get_tuple, methods=["POST"]),
+            Route("/list", list, methods=["POST"]),
+            Route("/put", put, methods=["POST"]),
+            Route("/put_writes", put_writes, methods=["POST"]),
+            Route("/ok", lambda _: ApiResponse({"ok": True}), methods=["GET"]),
+        ]
+    )
+
+    server = uvicorn.Server(
+        uvicorn.Config(
+            remote,
+            uds=CHECKPOINTER_SOCKET,
+            # We need to _explicitly_ set these values in order
+            # to avoid reinitialising the logger, which removes
+            # the structlog logger setup before.
+            # See: https://github.com/encode/uvicorn/blob/8f4c8a7f34914c16650ebd026127b96560425fde/uvicorn/config.py#L357-L393
+            log_config=None,
+            log_level=None,
+            access_log=True,
+        )
+    )
+    await server.serve()
+
+
+def _get_passthrough_store(conn: AsyncConnectionProto):
+    from langgraph_storage.store import Store
+
+    store = Store(conn)
+
+    return store
+
+
+async def run_remote_store():
+    from langgraph_storage.database import connect
+
+    async def abatch(request: Request):
+        payload = orjson.loads(await request.body())
+        operations = payload.get("operations", [])
+
+        if not operations:
+            return ApiResponse({"error": "No operations provided"}, status_code=400)
+
+        # Convert raw operations to proper objects
+        processed_operations = []
+        for op in operations:
+            if "value" in op:
+                processed_operations.append(
+                    PutOp(
+                        namespace=tuple(op["namespace"]),
+                        key=op["key"],
+                        value=op["value"],
+                    )
+                )
+            elif "namespace_prefix" in op:
+                processed_operations.append(
+                    SearchOp(
+                        namespace_prefix=tuple(op["namespace_prefix"]),
+                        filter=op.get("filter"),
+                        limit=op.get("limit", 10),
+                        offset=op.get("offset", 0),
+                    )
+                )
+
+            elif "namespace" in op and "key" in op:
+                processed_operations.append(
+                    GetOp(namespace=tuple(op["namespace"]), key=op["key"])
+                )
+            elif "match_conditions" in op:
+                processed_operations.append(
+                    ListNamespacesOp(
+                        match_conditions=tuple(op["match_conditions"]),
+                        max_depth=op.get("max_depth"),
+                        limit=op.get("limit", 100),
+                        offset=op.get("offset", 0),
+                    )
+                )
+            else:
+                return ApiResponse(
+                    {"error": f"Unknown operation type: {op}"}, status_code=400
+                )
+
+        async with connect() as conn:
+            store = _get_passthrough_store(conn)
+            results = await store.abatch(processed_operations)
+
+        # Handle potentially undefined or non-dict results
+        processed_results = []
+        # Result is of type: Union[Item, list[Item], list[tuple[str, ...]], None]
+        for result in results:
+            if isinstance(result, Item):
+                processed_results.append(result.dict())
+            elif isinstance(result, dict):
+                processed_results.append(result)
+            elif isinstance(result, list):
+                coerced = []
+                for res in result:
+                    if isinstance(res, Item):
+                        coerced.append(res.dict())
+                    elif isinstance(res, tuple):
+                        coerced.append(list(res))
+                    elif res is None:
+                        coerced.append(res)
+                    else:
+                        coerced.append(str(res))
+                processed_results.append(coerced)
+            elif result is None:
+                processed_results.append(None)
+            else:
+                processed_results.append(str(result))
+        return ApiResponse(processed_results)
+
+    # List all stores
+    async def aget(request: Request):
+        namespaces_str = request.query_params.get("namespaces")
+        key = request.query_params.get("key")
+
+        if not namespaces_str or not key:
+            return ApiResponse(
+                {"error": "Both namespaces and key are required"}, status_code=400
+            )
+
+        namespaces = namespaces_str.split(".")
+
+        async with connect() as conn:
+            store = _get_passthrough_store(conn)
+            result = await store.aget(namespaces, key)
+
+        return ApiResponse(result)
+
+    # Put the new store data
+    async def aput(request: Request):
+        payload = orjson.loads(await request.body())
+        namespace = tuple(payload["namespace"].split("."))
+        key = payload["key"]
+        value = payload["value"]
+
+        async with connect() as conn:
+            store = _get_passthrough_store(conn)
+            await store.aput(namespace, key, value)
+
+        return ApiResponse({"success": True})
+
+    # Search stores
+    async def asearch(request: Request):
+        payload = orjson.loads(await request.body())
+        namespace_prefix = tuple(payload["namespace_prefix"])
+        filter = payload.get("filter")
+        limit = payload.get("limit", 10)
+        offset = payload.get("offset", 0)
+
+        async with connect() as conn:
+            store = _get_passthrough_store(conn)
+            result = await store.asearch(
+                namespace_prefix, filter=filter, limit=limit, offset=offset
+            )
+
+        return ApiResponse([item.dict() for item in result])
+
+    # Delete store data
+    async def adelete(request: Request):
+        payload = orjson.loads(await request.body())
+        namespace = tuple(payload["namespace"])
+        key = payload["key"]
+
+        async with connect() as conn:
+            store = _get_passthrough_store(conn)
+            await store.adelete(namespace, key)
+
+        return ApiResponse({"success": True})
+
+    # List all namespaces
+    async def alist_namespaces(request: Request):
+        payload = orjson.loads(await request.body())
+        prefix = tuple(payload.get("prefix", [])) or None
+        suffix = tuple(payload.get("suffix", [])) or None
+        max_depth = payload.get("max_depth")
+        limit = payload.get("limit", 100)
+        offset = payload.get("offset", 0)
+
+        async with connect() as conn:
+            store = _get_passthrough_store(conn)
+            result = await store.alist_namespaces(
+                prefix=prefix,
+                suffix=suffix,
+                max_depth=max_depth,
+                limit=limit,
+                offset=offset,
+            )
+
+        return ApiResponse([list(ns) for ns in result])
+
+    remote = Starlette(
+        routes=[
+            Route("/items", aget, methods=["GET"]),
+            Route("/items", aput, methods=["PUT"]),
+            Route("/items", adelete, methods=["DELETE"]),
+            Route("/items/search", asearch, methods=["POST"]),
+            Route("/list/namespaces", alist_namespaces, methods=["POST"]),
+            Route("/items/batch", abatch, methods=["POST"]),
+            Route("/ok", lambda _: ApiResponse({"ok": True}), methods=["GET"]),
+        ]
+    )
+    server = uvicorn.Server(
+        uvicorn.Config(
+            remote,
+            uds=STORE_SOCKET,
+            # We need to _explicitly_ set these values in order
+            # to avoid reinitialising the logger, which removes
+            # the structlog logger setup before.
+            # See: https://github.com/encode/uvicorn/blob/8f4c8a7f34914c16650ebd026127b96560425fde/uvicorn/config.py#L357-L393
+            log_config=None,
+            log_level=None,
+            access_log=True,
+        )
+    )
+    await server.serve()
+
+
+async def wait_until_js_ready():
+    async with (
+        httpx.AsyncClient(
+            base_url="http://graph",
+            transport=httpx.AsyncHTTPTransport(uds=GRAPH_SOCKET),
+            limits=httpx.Limits(),
+        ) as graph_client,
+        httpx.AsyncClient(
+            base_url="http://checkpointer",
+            transport=httpx.AsyncHTTPTransport(uds=CHECKPOINTER_SOCKET),
+            limits=httpx.Limits(),
+        ) as checkpointer_client,
+        httpx.AsyncClient(
+            base_url="http://store",
+            transport=httpx.AsyncHTTPTransport(uds=STORE_SOCKET),
+            limits=httpx.Limits(),
+        ) as store_client,
+    ):
+        attempt = 0
+        while True:
+            try:
+                res = await graph_client.get("/ok")
+                res.raise_for_status()
+                res = await checkpointer_client.get("/ok")
+                res.raise_for_status()
+                res = await store_client.get("/ok")
+                res.raise_for_status()
+                return
+            except httpx.HTTPError:
+                if attempt > 240:
+                    raise
+                else:
+                    attempt += 1
+                    await asyncio.sleep(0.5)
+
+
+async def js_healthcheck():
+    async with (
+        httpx.AsyncClient(
+            base_url="http://graph",
+            transport=httpx.AsyncHTTPTransport(uds=GRAPH_SOCKET),
+            limits=httpx.Limits(),
+        ) as graph_client,
+        httpx.AsyncClient(
+            base_url="http://checkpointer",
+            transport=httpx.AsyncHTTPTransport(uds=CHECKPOINTER_SOCKET),
+            limits=httpx.Limits(),
+        ) as checkpointer_client,
+        httpx.AsyncClient(
+            base_url="http://store",
+            transport=httpx.AsyncHTTPTransport(uds=STORE_SOCKET),
+            limits=httpx.Limits(),
+        ) as store_client,
+    ):
+        try:
+            res = await graph_client.get("/ok")
+            res.raise_for_status()
+            res = await checkpointer_client.get("/ok")
+            res.raise_for_status()
+            res = await store_client.get("/ok")
+            res.raise_for_status()
+            return True
+        except httpx.HTTPError:
+            return False