ommlds 0.0.0.dev456__py3-none-any.whl → 0.0.0.dev485__py3-none-any.whl

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

@@ -22,9 +22,15 @@ 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')
 
@@ -34,13 +40,14 @@ class OpenaiEmbeddingService:
             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

@@ -2,31 +2,33 @@ 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 AiChat
 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 ToolUseMessage
 from ....chat.messages import ToolUseResultMessage
 from ....chat.messages import UserMessage
-from ....chat.stream.types import ContentAiChoiceDelta
+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.jsonschema import build_tool_spec_params_json_schema
 from ....tools.types import ToolSpec
 from ....tools.types import ToolUse
 from ....types import Option
@@ -35,61 +37,115 @@ from ....types import Option
 ##
 
 
-def build_request_messages(chat: Chat) -> ta.Sequence[ta.Mapping[str, ta.Any]]:
-    out: list[dict[str, ta.Any]] = []
+def build_oai_request_msgs(mc_chat: Chat) -> ta.Sequence[pt.ChatCompletionMessage]:
+    oai_msgs: list[pt.ChatCompletionMessage] = []
 
-    for m in chat:
-        if isinstance(m, SystemMessage):
-            out.append(dict(
-                role='system',
-                content=m.c,
+    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):
-            out.append(dict(
-                role='assistant',
-                content=check.isinstance(m.c, (str, None)),
+        elif isinstance(mc_msg, AiMessage):
+            oai_msgs.append(pt.AssistantChatCompletionMessage(
+                content=check.isinstance(mc_msg.c, (str, None)),
             ))
 
-        elif isinstance(m, ToolUseMessage):
-            out.append(dict(
-                role='assistant',
-                tool_calls=[
-                    dict(
-                        id=m.tu.id,
-                        function=dict(
-                            arguments=check.not_none(m.tu.raw_args),
-                            name=m.tu.name,
-                        ),
-                        type='function',
+        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,
                     ),
-                ],
+                )],
             ))
 
-        elif isinstance(m, UserMessage):
-            out.append(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(m, ToolUseResultMessage):
+        elif isinstance(mc_msg, ToolUseResultMessage):
             tc: str
-            if isinstance(m.tur.c, str):
-                tc = m.tur.c
-            elif isinstance(m.tur.c, JsonContent):
-                tc = json.dumps_compact(m.tur.c)
+            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(m.tur.c)
-            out.append(dict(
-                role='tool',
-                tool_call_id=m.tur.id,
+                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(m)
+            raise TypeError(mc_msg)
 
-    return out
+    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)
+
+
+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 ValueError(delta)
 
 
 ##
@@ -110,22 +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',
-        ToolUseMessage: 'assistant',
-        ToolUseResultMessage: 'tool',
-    }
-
     DEFAULT_OPTIONS: ta.ClassVar[tv.TypedValues[Option]] = tv.TypedValues[Option](
-        Temperature(0.),
-        MaxTokens(1024),
+        # 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):
@@ -135,8 +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] = {}
@@ -160,69 +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_messages(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_chat(self, message: ta.Mapping[str, ta.Any]) -> AiChat:
-        out: list[AnyAiMessage] = []
-        if (c := message.get('content')) is not None:
-            out.append(AiMessage(c))
-        for tc in message.get('tool_calls', []):
-            out.append(ToolUseMessage(
-                ToolUse(
-                    id=tc['id'],
-                    name=tc['function']['name'],
-                    args=json.loads(tc['function']['arguments'] or '{}'),
-                    raw_args=tc['function']['arguments'],
-                ),
-            ))
-        return out
-
-    def build_response(self, raw_response: ta.Mapping[str, ta.Any]) -> ChatChoicesResponse:
-        return ChatChoicesResponse(
-            [
-                AiChoice(self.build_ai_chat(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_choice_delta(self, delta: ta.Mapping[str, ta.Any]) -> ContentAiChoiceDelta:
-        return ContentAiChoiceDelta(
-            check.not_none(delta.get('content')),
-            # FIXME:
-            # tool_exec_requests=[
-            #     ToolUse(
-            #         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,
-        )

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

@@ -30,10 +30,12 @@ _GPT_MODEL_NAMES = [
     'gpt-5-chat-latest',
     'gpt-5-mini',
     'gpt-5-nano',
+
+    'gpt-5.1',
 ]
 
 
-MODEL_NAMES = ModelNameCollection(
+CHAT_MODEL_NAMES = ModelNameCollection(
     default='gpt',
     aliases={
         **{
@@ -46,8 +48,8 @@ MODEL_NAMES = ModelNameCollection(
             for n in _GPT_MODEL_NAMES
         },
 
-        'gpt': 'gpt-4o',
-        'gpt-mini': 'gpt-4o-mini',
+        'gpt': 'gpt-5.1',
+        'gpt-mini': 'gpt-5-mini',
 
         #
 
@@ -61,11 +63,35 @@ MODEL_NAMES = ModelNameCollection(
 
 
 # @omlish-manifest
-_BACKEND_STRINGS_MANIFEST = BackendStringsManifest(
+_CHAT_BACKEND_STRINGS_MANIFEST = BackendStringsManifest(
     [
         'ChatChoicesService',
         'ChatChoicesStreamService',
     ],
     'openai',
-    model_names=MODEL_NAMES,
+    model_names=CHAT_MODEL_NAMES,
+)
+
+
+##
+
+
+# @omlish-manifest
+_COMPLETION_BACKEND_STRINGS_MANIFEST = BackendStringsManifest(
+    [
+        'CompletionService',
+    ],
+    'openai',
+)
+
+
+##
+
+
+# @omlish-manifest
+_EMBEDDING_BACKEND_STRINGS_MANIFEST = BackendStringsManifest(
+    [
+        'EmbeddingService',
+    ],
+    'openai',
 )

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

@@ -11,14 +11,14 @@ from omlish.http import all as http
 from omlish.http import sse
 from omlish.io.buffers import DelimitingBuffer
 
-from .....backends.openai.protocol.chatcompletion.chunk import ChatCompletionChunk
+from .....backends.openai import protocol as pt
 from ....chat.choices.services import ChatChoicesOutputs
-from ....chat.stream.services import ChatChoicesStreamRequest
-from ....chat.stream.services import ChatChoicesStreamResponse
-from ....chat.stream.services import static_check_is_chat_choices_stream_service
-from ....chat.stream.types import AiChoiceDeltas
-from ....chat.stream.types import AiChoicesDeltas
-from ....chat.stream.types import ChatChoicesStreamOption
+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.stream.types import ChatChoicesStreamOption
 from ....configs import Config
 from ....resources import ResourcesOption
 from ....resources import UseResources
@@ -28,7 +28,8 @@ from ....stream.services import StreamResponseSink
 from ....stream.services import new_stream_response
 from .chat import OpenaiChatChoicesService
 from .format import OpenaiChatRequestHandler
-from .names import MODEL_NAMES
+from .format import build_mc_ai_delta
+from .names import CHAT_MODEL_NAMES
 
 
 ##
@@ -40,14 +41,20 @@ from .names import MODEL_NAMES
 # )
 @static_check_is_chat_choices_stream_service
 class OpenaiChatChoicesStreamService:
-    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._model_name = cc.pop(OpenaiChatChoicesService.DEFAULT_MODEL_NAME)
             self._api_key = ApiKey.pop_secret(cc, env='OPENAI_API_KEY')
 
-    READ_CHUNK_SIZE = 64 * 1024
+    READ_CHUNK_SIZE: ta.ClassVar[int] = -1
 
     async def invoke(self, request: ChatChoicesStreamRequest) -> ChatChoicesStreamResponse:
         # check.isinstance(request, ChatRequest)
@@ -59,16 +66,16 @@ class OpenaiChatChoicesStreamService:
                 for o in request.options
                 if not isinstance(o, (ChatChoicesStreamOption, StreamOption, ResourcesOption))
             ],
-            model=MODEL_NAMES.resolve(self._model_name.v),
+            model=CHAT_MODEL_NAMES.resolve(self._model_name.v),
             mandatory_kwargs=dict(
                 stream=True,
-                stream_options=dict(
+                stream_options=pt.ChatCompletionRequest.StreamOptions(
                     include_usage=True,
                 ),
             ),
         )
 
-        raw_request = rh.raw_request()
+        raw_request = msh.marshal(rh.oai_request())
 
         http_request = http.HttpRequest(
             'https://api.openai.com/v1/chat/completions',
@@ -80,15 +87,14 @@ class OpenaiChatChoicesStreamService:
         )
 
         async with UseResources.or_new(request.options) as rs:
-            http_client = rs.enter_context(http.client())
-            http_response = rs.enter_context(http_client.stream_request(http_request))
+            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]:
                 db = DelimitingBuffer([b'\r', b'\n', b'\r\n'])
                 sd = sse.SseDecoder()
                 while True:
-                    # FIXME: read1 not on response stream protocol
-                    b = http_response.stream.read1(self.READ_CHUNK_SIZE)  # type: ignore[attr-defined]
+                    b = await http_response.stream.read1(self.READ_CHUNK_SIZE)
                     for l in db.feed(b):
                         if isinstance(l, DelimitingBuffer.Incomplete):
                             # FIXME: handle
@@ -105,20 +111,20 @@ class OpenaiChatChoicesStreamService:
 
                                 check.state(sj['object'] == 'chat.completion.chunk')
 
-                                ccc = msh.unmarshal(sj, ChatCompletionChunk)  # noqa
-                                # print(ccc)
+                                ccc = msh.unmarshal(sj, pt.ChatCompletionChunk)
 
                                 # FIXME: stop reason
-                                if not sj['choices']:
+                                if not ccc.choices:
                                     continue
 
-                                if any(choice['delta'] for choice in sj['choices']):
-                                    await sink.emit(AiChoicesDeltas([
-                                        AiChoiceDeltas(
-                                            [rh.build_ai_choice_delta(choice['delta'])] if choice['delta'] else [],
-                                        )
-                                        for choice in sj['choices']
-                                    ]))
+                                if any(choice.finish_reason for choice in ccc.choices):
+                                    check.state(all(choice.finish_reason for choice in ccc.choices))
+                                    break
+
+                                await sink.emit(AiChoicesDeltas([
+                                    AiChoiceDeltas([build_mc_ai_delta(choice.delta)])
+                                    for choice in ccc.choices
+                                ]))
 
                     if not b:
                         return []

ommlds/minichain/backends/impls/sentencepiece/tokens.py

@@ -1,23 +1,26 @@
 import typing as ta
 
-import sentencepiece as spm
-
 from omlish import check
+from omlish import lang
 
 from .... import tokens as tks
 
 
+with lang.auto_proxy_import(globals()):
+    import sentencepiece as spm
+
+
 ##
 
 
-def build_vocab(spm_tokenizer: spm.SentencePieceProcessor) -> tks.Vocab:
+def build_vocab(spm_tokenizer: 'spm.SentencePieceProcessor') -> tks.Vocab:
     return tks.Vocab([
         (ta.cast(tks.Token, i), tks.TokenStr(spm_tokenizer.id_to_piece(i)))  # noqa
         for i in range(spm_tokenizer.get_piece_size())  # noqa
     ])
 
 
-def build_specials(spm_tokenizer: spm.SentencePieceProcessor) -> tks.SpecialTokens:
+def build_specials(spm_tokenizer: 'spm.SentencePieceProcessor') -> tks.SpecialTokens:
     # FIXME
     return tks.SpecialTokens([])
 
@@ -28,7 +31,7 @@ def build_specials(spm_tokenizer: spm.SentencePieceProcessor) -> tks.SpecialToke
 class SentencepieceTokenizer(tks.BaseTokenizer):
     def __init__(
             self,
-            spm_tokenizer: spm.SentencePieceProcessor,
+            spm_tokenizer: 'spm.SentencePieceProcessor',
     ) -> None:
         self._spm_tokenizer = check.isinstance(spm_tokenizer, spm.SentencePieceProcessor)
 
@@ -38,7 +41,7 @@ class SentencepieceTokenizer(tks.BaseTokenizer):
         )
 
     @property
-    def spm_tokenizer(self) -> spm.SentencePieceProcessor:
+    def spm_tokenizer(self) -> 'spm.SentencePieceProcessor':
         return self._spm_tokenizer
 
     #

ommlds/minichain/backends/impls/tavily.py

@@ -0,0 +1,66 @@
+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.tavily import protocol as pt
+from ...search import SearchHit
+from ...search import SearchHits
+from ...search import SearchRequest
+from ...search import SearchResponse
+from ...search import static_check_is_search_service
+from ...standard import ApiKey
+
+
+##
+
+
+# @omlish-manifest $.minichain.registries.manifests.RegistryManifest(
+#     name='tavily',
+#     type='SearchService',
+# )
+@static_check_is_search_service
+class TavilySearchService:
+    def __init__(
+            self,
+            *configs: ApiKey,
+            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='TAVILY_API_KEY')
+
+    async def invoke(self, request: SearchRequest) -> SearchResponse:
+        pt_request = pt.SearchRequest(
+            query=request.v,
+        )
+
+        raw_request = msh.marshal(pt_request)
+
+        http_response = await http.async_request(
+            'https://api.tavily.com/search',
+            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'))
+
+        pt_response = msh.unmarshal(raw_response, pt.SearchResponse)
+
+        return SearchResponse(SearchHits(
+            l=[
+                SearchHit(
+                    title=r.title,
+                    url=r.url,
+                )
+                for r in pt_response.results or []
+            ],
+        ))