ommlds 0.0.0.dev436__py3-none-any.whl → 0.0.0.dev480__py3-none-any.whl

ommlds/minichain/backends/impls/ollama/chat.py

@@ -0,0 +1,199 @@
+import typing as ta
+
+from omlish import check
+from omlish import lang
+from omlish import marshal as msh
+from omlish import typedvalues as tv
+from omlish.formats import json
+from omlish.http import all as http
+from omlish.io.buffers import DelimitingBuffer
+
+from .....backends.ollama import protocol as pt
+from ....chat.choices.services import ChatChoicesOutputs
+from ....chat.choices.services import ChatChoicesRequest
+from ....chat.choices.services import ChatChoicesResponse
+from ....chat.choices.services import static_check_is_chat_choices_service
+from ....chat.choices.stream.services import ChatChoicesStreamRequest
+from ....chat.choices.stream.services import ChatChoicesStreamResponse
+from ....chat.choices.stream.services import static_check_is_chat_choices_stream_service
+from ....chat.choices.stream.types import AiChoiceDeltas
+from ....chat.choices.stream.types import AiChoicesDeltas
+from ....chat.choices.types import AiChoice
+from ....chat.messages import AiMessage
+from ....chat.messages import AnyAiMessage
+from ....chat.messages import Message
+from ....chat.messages import SystemMessage
+from ....chat.messages import UserMessage
+from ....chat.stream.types import ContentAiDelta
+from ....models.configs import ModelName
+from ....resources import UseResources
+from ....standard import ApiUrl
+from ....stream.services import StreamResponseSink
+from ....stream.services import new_stream_response
+
+
+##
+
+
+# @omlish-manifest $.minichain.backends.strings.manifests.BackendStringsManifest(
+#     [
+#         'ChatChoicesService',
+#         'ChatChoicesStreamService',
+#     ],
+#     'ollama',
+# )
+
+
+##
+
+
+class BaseOllamaChatChoicesService(lang.Abstract):
+    DEFAULT_API_URL: ta.ClassVar[ApiUrl] = ApiUrl('http://localhost:11434/api')
+    DEFAULT_MODEL_NAME: ta.ClassVar[ModelName] = ModelName('llama3.2')
+
+    def __init__(
+            self,
+            *configs: ApiUrl | ModelName,
+            http_client: http.AsyncHttpClient | None = None,
+    ) -> None:
+        super().__init__()
+
+        self._http_client = http_client
+
+        with tv.consume(*configs) as cc:
+            self._api_url = cc.pop(self.DEFAULT_API_URL)
+            self._model_name = cc.pop(self.DEFAULT_MODEL_NAME)
+
+    #
+
+    ROLE_MAP: ta.ClassVar[ta.Mapping[type[Message], pt.Role]] = {  # noqa
+        SystemMessage: 'system',
+        UserMessage: 'user',
+        AiMessage: 'assistant',
+    }
+
+    @classmethod
+    def _get_message_content(cls, m: Message) -> str | None:
+        if isinstance(m, (AiMessage, UserMessage, SystemMessage)):
+            return check.isinstance(m.c, str)
+        else:
+            raise TypeError(m)
+
+    @classmethod
+    def _build_request_messages(cls, mc_msgs: ta.Iterable[Message]) -> ta.Sequence[pt.Message]:
+        messages: list[pt.Message] = []
+        for m in mc_msgs:
+            messages.append(pt.Message(
+                role=cls.ROLE_MAP[type(m)],
+                content=cls._get_message_content(m),
+            ))
+        return messages
+
+
+##
+
+
+# @omlish-manifest $.minichain.registries.manifests.RegistryManifest(
+#     name='ollama',
+#     type='ChatChoicesService',
+# )
+@static_check_is_chat_choices_service
+class OllamaChatChoicesService(BaseOllamaChatChoicesService):
+    async def invoke(
+            self,
+            request: ChatChoicesRequest,
+    ) -> ChatChoicesResponse:
+        messages = self._build_request_messages(request.v)
+
+        a_req = pt.ChatRequest(
+            model=self._model_name.v,
+            messages=messages,
+            # tools=tools or None,
+            stream=False,
+        )
+
+        raw_request = msh.marshal(a_req)
+
+        async with http.manage_async_client(self._http_client) as http_client:
+            raw_response = await http_client.request(http.HttpRequest(
+                self._api_url.v.removesuffix('/') + '/chat',
+                data=json.dumps(raw_request).encode('utf-8'),
+            ))
+
+        json_response = json.loads(check.not_none(raw_response.data).decode('utf-8'))
+
+        resp = msh.unmarshal(json_response, pt.ChatResponse)
+
+        out: list[AnyAiMessage] = []
+        if resp.message.role == 'assistant':
+            out.append(AiMessage(
+                check.not_none(resp.message.content),
+            ))
+        else:
+            raise TypeError(resp.message.role)
+
+        return ChatChoicesResponse([
+            AiChoice(out),
+        ])
+
+
+##
+
+
+# @omlish-manifest $.minichain.registries.manifests.RegistryManifest(
+#     name='ollama',
+#     type='ChatChoicesStreamService',
+# )
+@static_check_is_chat_choices_stream_service
+class OllamaChatChoicesStreamService(BaseOllamaChatChoicesService):
+    READ_CHUNK_SIZE: ta.ClassVar[int] = -1
+
+    async def invoke(
+            self,
+            request: ChatChoicesStreamRequest,
+    ) -> ChatChoicesStreamResponse:
+        messages = self._build_request_messages(request.v)
+
+        a_req = pt.ChatRequest(
+            model=self._model_name.v,
+            messages=messages,
+            # tools=tools or None,
+            stream=True,
+        )
+
+        raw_request = msh.marshal(a_req)
+
+        http_request = http.HttpRequest(
+            self._api_url.v.removesuffix('/') + '/chat',
+            data=json.dumps(raw_request).encode('utf-8'),
+        )
+
+        async with UseResources.or_new(request.options) as rs:
+            http_client = await rs.enter_async_context(http.manage_async_client(self._http_client))
+            http_response = await rs.enter_async_context(await http_client.stream_request(http_request))
+
+            async def inner(sink: StreamResponseSink[AiChoicesDeltas]) -> ta.Sequence[ChatChoicesOutputs] | None:
+                db = DelimitingBuffer([b'\r', b'\n', b'\r\n'])
+                while True:
+                    b = await http_response.stream.read1(self.READ_CHUNK_SIZE)
+                    for l in db.feed(b):
+                        if isinstance(l, DelimitingBuffer.Incomplete):
+                            # FIXME: handle
+                            return []
+
+                        lj = json.loads(l.decode('utf-8'))
+                        lp: pt.ChatResponse = msh.unmarshal(lj, pt.ChatResponse)
+
+                        check.state(lp.message.role == 'assistant')
+                        check.none(lp.message.tool_name)
+                        check.state(not lp.message.tool_calls)
+
+                        if (c := lp.message.content):
+                            await sink.emit(AiChoicesDeltas([AiChoiceDeltas([ContentAiDelta(
+                                c,
+                            )])]))
+
+                    if not b:
+                        return []
+
+            return await new_stream_response(rs, inner)

ommlds/minichain/backends/impls/openai/chat.py

@@ -14,10 +14,12 @@ TODO:
 import typing as ta
 
 from omlish import check
+from omlish import marshal as msh
 from omlish import typedvalues as tv
 from omlish.formats import json
 from omlish.http import all as http
 
+from .....backends.openai import protocol as pt
 from ....chat.choices.services import ChatChoicesRequest
 from ....chat.choices.services import ChatChoicesResponse
 from ....chat.choices.services import static_check_is_chat_choices_service
@@ -25,7 +27,8 @@ from ....models.configs import ModelName
 from ....standard import ApiKey
 from ....standard import DefaultOptions
 from .format import OpenaiChatRequestHandler
-from .names import MODEL_NAMES
+from .format import build_mc_choices_response
+from .names import CHAT_MODEL_NAMES
 
 
 ##
@@ -37,17 +40,23 @@ from .names import MODEL_NAMES
 # )
 @static_check_is_chat_choices_service
 class OpenaiChatChoicesService:
-    DEFAULT_MODEL_NAME: ta.ClassVar[ModelName] = ModelName(check.not_none(MODEL_NAMES.default))
+    DEFAULT_MODEL_NAME: ta.ClassVar[ModelName] = ModelName(check.not_none(CHAT_MODEL_NAMES.default))
 
-    def __init__(self, *configs: ApiKey | ModelName | DefaultOptions) -> None:
+    def __init__(
+            self,
+            *configs: ApiKey | ModelName | DefaultOptions,
+            http_client: http.AsyncHttpClient | None = None,
+    ) -> None:
         super().__init__()
 
+        self._http_client = http_client
+
         with tv.consume(*configs) as cc:
             self._model_name = cc.pop(self.DEFAULT_MODEL_NAME)
             self._api_key = ApiKey.pop_secret(cc, env='OPENAI_API_KEY')
             self._default_options: tv.TypedValues = DefaultOptions.pop(cc)
 
-    def invoke(self, request: ChatChoicesRequest) -> ChatChoicesResponse:
+    async def invoke(self, request: ChatChoicesRequest) -> ChatChoicesResponse:
         # check.isinstance(request, ChatRequest)
 
         rh = OpenaiChatRequestHandler(
@@ -57,23 +66,24 @@ class OpenaiChatChoicesService:
                 *request.options,
                 override=True,
             ),
-            model=MODEL_NAMES.resolve(self._model_name.v),
+            model=CHAT_MODEL_NAMES.resolve(self._model_name.v),
             mandatory_kwargs=dict(
                 stream=False,
             ),
         )
 
-        raw_request = rh.raw_request()
+        raw_request = msh.marshal(rh.oai_request())
 
-        http_response = http.request(
+        http_response = await http.async_request(
             'https://api.openai.com/v1/chat/completions',
             headers={
                 http.consts.HEADER_CONTENT_TYPE: http.consts.CONTENT_TYPE_JSON,
                 http.consts.HEADER_AUTH: http.consts.format_bearer_auth_header(check.not_none(self._api_key).reveal()),
             },
             data=json.dumps(raw_request).encode('utf-8'),
+            client=self._http_client,
         )
 
         raw_response = json.loads(check.not_none(http_response.data).decode('utf-8'))
 
-        return rh.build_response(raw_response)
+        return build_mc_choices_response(msh.unmarshal(raw_response, pt.ChatCompletionResponse))

ommlds/minichain/backends/impls/openai/completion.py

@@ -23,13 +23,19 @@ from ....standard import ApiKey
 class OpenaiCompletionService:
     DEFAULT_MODEL_NAME: ta.ClassVar[str] = 'gpt-3.5-turbo-instruct'
 
-    def __init__(self, *configs: Config) -> None:
+    def __init__(
+            self,
+            *configs: Config,
+            http_client: http.AsyncHttpClient | None = None,
+    ) -> None:
         super().__init__()
 
+        self._http_client = http_client
+
         with tv.consume(*configs) as cc:
             self._api_key = ApiKey.pop_secret(cc, env='OPENAI_API_KEY')
 
-    def invoke(self, t: CompletionRequest) -> CompletionResponse:
+    async def invoke(self, t: CompletionRequest) -> CompletionResponse:
         raw_request = dict(
             model=self.DEFAULT_MODEL_NAME,
             prompt=t.v,
@@ -41,13 +47,14 @@ class OpenaiCompletionService:
             stream=False,
         )
 
-        raw_response = http.request(
+        raw_response = await http.async_request(
             'https://api.openai.com/v1/completions',
             headers={
                 http.consts.HEADER_CONTENT_TYPE: http.consts.CONTENT_TYPE_JSON,
                 http.consts.HEADER_AUTH: http.consts.format_bearer_auth_header(check.not_none(self._api_key).reveal()),
             },
             data=json.dumps(raw_request).encode('utf-8'),
+            client=self._http_client,
         )
 
         response = json.loads(check.not_none(raw_response.data).decode('utf-8'))

ommlds/minichain/backends/impls/openai/embedding.py

@@ -22,25 +22,32 @@ from ....vectors.types import Vector
 class OpenaiEmbeddingService:
     model = 'text-embedding-3-small'
 
-    def __init__(self, *configs: Config) -> None:
+    def __init__(
+            self,
+            *configs: Config,
+            http_client: http.AsyncHttpClient | None = None,
+    ) -> None:
         super().__init__()
 
+        self._http_client = http_client
+
         with tv.consume(*configs) as cc:
             self._api_key = ApiKey.pop_secret(cc, env='OPENAI_API_KEY')
 
-    def invoke(self, request: EmbeddingRequest) -> EmbeddingResponse:
+    async def invoke(self, request: EmbeddingRequest) -> EmbeddingResponse:
         raw_request = dict(
             model=self.model,
             input=check.isinstance(request.v, str),
         )
 
-        raw_response = http.request(
+        raw_response = await http.async_request(
             'https://api.openai.com/v1/embeddings',
             headers={
                 http.consts.HEADER_CONTENT_TYPE: http.consts.CONTENT_TYPE_JSON,
                 http.consts.HEADER_AUTH: http.consts.format_bearer_auth_header(check.not_none(self._api_key).reveal()),
             },
             data=json.dumps(raw_request).encode('utf-8'),
+            client=self._http_client,
         )
 
         response = json.loads(check.not_none(raw_response.data).decode('utf-8'))

ommlds/minichain/backends/impls/openai/format.py

@@ -2,74 +2,150 @@ import typing as ta
 
 from omlish import cached
 from omlish import check
-from omlish import lang
 from omlish import typedvalues as tv
 from omlish.formats import json
 
+from .....backends.openai import protocol as pt
 from ....chat.choices.services import ChatChoicesResponse
 from ....chat.choices.types import AiChoice
+from ....chat.choices.types import AiChoices
 from ....chat.choices.types import ChatChoicesOptions
 from ....chat.messages import AiMessage
+from ....chat.messages import AnyAiMessage
 from ....chat.messages import Chat
-from ....chat.messages import Message
 from ....chat.messages import SystemMessage
-from ....chat.messages import ToolExecResultMessage
+from ....chat.messages import ToolUseMessage
+from ....chat.messages import ToolUseResultMessage
 from ....chat.messages import UserMessage
-from ....chat.stream.types import AiMessageDelta
+from ....chat.stream.types import AiDelta
+from ....chat.stream.types import ContentAiDelta
+from ....chat.stream.types import PartialToolUseAiDelta
 from ....chat.tools.types import Tool
+from ....content.json import JsonContent
 from ....content.prepare import prepare_content_str
+from ....llms.types import MaxCompletionTokens
 from ....llms.types import MaxTokens
 from ....llms.types import Temperature
 from ....llms.types import TokenUsage
 from ....llms.types import TokenUsageOutput
-from ....tools.jsonschema import build_tool_spec_json_schema
-from ....tools.types import ToolExecRequest
+from ....tools.jsonschema import build_tool_spec_params_json_schema
 from ....tools.types import ToolSpec
+from ....tools.types import ToolUse
 from ....types import Option
 
 
 ##
 
 
-def build_request_message(m: Message) -> ta.Mapping[str, ta.Any]:
-    if isinstance(m, SystemMessage):
-        return dict(
-            role='system',
-            content=m.c,
-        )
+def build_oai_request_msgs(mc_chat: Chat) -> ta.Sequence[pt.ChatCompletionMessage]:
+    oai_msgs: list[pt.ChatCompletionMessage] = []
+
+    for mc_msg in mc_chat:
+        if isinstance(mc_msg, SystemMessage):
+            oai_msgs.append(pt.SystemChatCompletionMessage(
+                content=check.isinstance(mc_msg.c, str),
+            ))
 
-    elif isinstance(m, AiMessage):
-        return dict(
-            role='assistant',
-            content=check.isinstance(m.c, (str, None)),
-            **(dict(tool_calls=[
-                dict(
-                    id=te.id,
-                    function=dict(
-                        arguments=check.not_none(te.raw_args),
-                        name=te.name,
+        elif isinstance(mc_msg, AiMessage):
+            oai_msgs.append(pt.AssistantChatCompletionMessage(
+                content=check.isinstance(mc_msg.c, (str, None)),
+            ))
+
+        elif isinstance(mc_msg, ToolUseMessage):
+            oai_msgs.append(pt.AssistantChatCompletionMessage(
+                tool_calls=[pt.AssistantChatCompletionMessage.ToolCall(
+                    id=check.not_none(mc_msg.tu.id),
+                    function=pt.AssistantChatCompletionMessage.ToolCall.Function(
+                        arguments=check.not_none(mc_msg.tu.raw_args),
+                        name=mc_msg.tu.name,
                     ),
-                    type='function',
-                )
-                for te in m.tool_exec_requests
-            ]) if m.tool_exec_requests else {}),
-        )
+                )],
+            ))
 
-    elif isinstance(m, UserMessage):
-        return dict(
-            role='user',
-            content=prepare_content_str(m.c),
-        )
+        elif isinstance(mc_msg, UserMessage):
+            oai_msgs.append(pt.UserChatCompletionMessage(
+                content=prepare_content_str(mc_msg.c),
+            ))
+
+        elif isinstance(mc_msg, ToolUseResultMessage):
+            tc: str
+            if isinstance(mc_msg.tur.c, str):
+                tc = mc_msg.tur.c
+            elif isinstance(mc_msg.tur.c, JsonContent):
+                tc = json.dumps_compact(mc_msg.tur.c)
+            else:
+                raise TypeError(mc_msg.tur.c)
+            oai_msgs.append(pt.ToolChatCompletionMessage(
+                tool_call_id=check.not_none(mc_msg.tur.id),
+                content=tc,
+            ))
+
+        else:
+            raise TypeError(mc_msg)
+
+    return oai_msgs
+
+
+#
+
+
+def build_mc_ai_choice(oai_choice: pt.ChatCompletionResponseChoice) -> AiChoice:
+    cur: list[AnyAiMessage] = []
+
+    oai_msg = oai_choice.message
+
+    if (oai_c := oai_msg.content) is not None:
+        cur.append(AiMessage(check.isinstance(oai_c, str)))
+
+    for oai_tc in oai_msg.tool_calls or []:
+        cur.append(ToolUseMessage(ToolUse(
+            id=oai_tc.id,
+            name=oai_tc.function.name,
+            args=json.loads(oai_tc.function.arguments or '{}'),
+            raw_args=oai_tc.function.arguments,
+        )))
+
+    return AiChoice(cur)
 
-    elif isinstance(m, ToolExecResultMessage):
-        return dict(
-            role='tool',
-            tool_call_id=m.id,
-            content=check.isinstance(m.c, str),
+
+def build_mc_ai_choices(oai_resp: pt.ChatCompletionResponse) -> AiChoices:
+    return [
+        build_mc_ai_choice(oai_choice)
+        for oai_choice in oai_resp.choices
+    ]
+
+
+def build_mc_choices_response(oai_resp: pt.ChatCompletionResponse) -> ChatChoicesResponse:
+    return ChatChoicesResponse(
+        build_mc_ai_choices(oai_resp),
+
+        tv.TypedValues(
+            *([TokenUsageOutput(TokenUsage(
+                input=tu.prompt_tokens,
+                output=tu.completion_tokens,
+                total=tu.total_tokens,
+            ))] if (tu := oai_resp.usage) is not None else []),
+        ),
+    )
+
+
+def build_mc_ai_delta(delta: pt.ChatCompletionChunkChoiceDelta) -> AiDelta:
+    if delta.content is not None:
+        check.state(not delta.tool_calls)
+        return ContentAiDelta(delta.content)
+
+    elif delta.tool_calls is not None:
+        check.state(delta.content is None)
+        tc = check.single(delta.tool_calls)
+        tc_fn = check.not_none(tc.function)
+        return PartialToolUseAiDelta(
+            id=tc.id,
+            name=tc_fn.name,
+            raw_args=tc_fn.arguments,
         )
 
     else:
-        raise TypeError(m)
+        raise ValueError(delta)
 
 
 ##
@@ -90,21 +166,15 @@ class OpenaiChatRequestHandler:
         self._model = model
         self._mandatory_kwargs = mandatory_kwargs
 
-    ROLES_MAP: ta.ClassVar[ta.Mapping[type[Message], str]] = {
-        SystemMessage: 'system',
-        UserMessage: 'user',
-        AiMessage: 'assistant',
-        ToolExecResultMessage: 'tool',
-    }
-
-    DEFAULT_OPTIONS: ta.ClassVar[tv.TypedValues[Option]] = tv.TypedValues(
-        Temperature(0.),
-        MaxTokens(1024),
+    DEFAULT_OPTIONS: ta.ClassVar[tv.TypedValues[Option]] = tv.TypedValues[Option](
+        # Temperature(0.),
+        # MaxTokens(1024),
     )
 
     _OPTION_KWARG_NAMES_MAP: ta.ClassVar[ta.Mapping[str, type[ChatChoicesOptions]]] = dict(
         temperature=Temperature,
         max_tokens=MaxTokens,
+        max_completion_tokens=MaxCompletionTokens,
     )
 
     class _ProcessedOptions(ta.NamedTuple):
@@ -114,8 +184,8 @@ class OpenaiChatRequestHandler:
     @cached.function
     def _process_options(self) -> _ProcessedOptions:
         kwargs: dict = dict(
-            temperature=0,
-            max_tokens=1024,
+            # temperature=0,
+            # max_tokens=1024,
         )
 
         tools_by_name: dict[str, ToolSpec] = {}
@@ -139,71 +209,26 @@ class OpenaiChatRequestHandler:
         )
 
     @cached.function
-    def raw_request(self) -> ta.Mapping[str, ta.Any]:
+    def oai_request(self) -> pt.ChatCompletionRequest:
         po = self._process_options()
 
-        tools = [
-            dict(
-                type='function',
-                function=build_tool_spec_json_schema(ts),
+        tools: list[pt.ChatCompletionRequestTool] = [
+            pt.ChatCompletionRequestTool(
+                function=pt.ChatCompletionRequestTool.Function(
+                    name=check.not_none(ts.name),
+                    description=prepare_content_str(ts.desc),
+                    parameters=build_tool_spec_params_json_schema(ts),
+                ),
             )
             for ts in po.tools_by_name.values()
         ]
 
-        return dict(
+        return pt.ChatCompletionRequest(
             model=self._model,
-            messages=[
-                build_request_message(m)
-                for m in self._chat
-            ],
+            messages=build_oai_request_msgs(self._chat),
             top_p=1,
-            **lang.opt_kw(tools=tools),
+            tools=tools or None,
             frequency_penalty=0.0,
             presence_penalty=0.0,
             **po.kwargs,
         )
-
-    def build_ai_message(self, message: ta.Mapping[str, ta.Any]) -> AiMessage:
-        return AiMessage(
-            message.get('content'),
-            tool_exec_requests=[
-                ToolExecRequest(
-                    id=tc['id'],
-                    name=tc['function']['name'],
-                    args=json.loads(tc['function']['arguments'] or '{}'),
-                    raw_args=tc['function']['arguments'],
-                )
-                for tc in message.get('tool_calls', [])
-            ] or None,
-        )
-
-    def build_response(self, raw_response: ta.Mapping[str, ta.Any]) -> ChatChoicesResponse:
-        return ChatChoicesResponse(
-            [
-                AiChoice(self.build_ai_message(choice['message']))
-                for choice in raw_response['choices']
-            ],
-
-            tv.TypedValues(
-                *([TokenUsageOutput(TokenUsage(
-                    input=tu['prompt_tokens'],
-                    output=tu['completion_tokens'],
-                    total=tu['total_tokens'],
-                ))] if (tu := raw_response.get('usage')) is not None else []),
-            ),
-        )
-
-    def build_ai_message_delta(self, delta: ta.Mapping[str, ta.Any]) -> AiMessageDelta:
-        return AiMessageDelta(
-            delta.get('content'),
-            # FIXME:
-            # tool_exec_requests=[
-            #     ToolExecRequest(
-            #         id=tc['id'],
-            #         spec=self._process_options().tools_by_name[tc['function']['name']],
-            #         args=json.loads(tc['function']['arguments'] or '{}'),
-            #         raw_args=tc['function']['arguments'],
-            #     )
-            #     for tc in message_or_delta.get('tool_calls', [])
-            # ] or None,
-        )