ai-microcore 4.0.0.dev3__tar.gz → 4.0.0.dev5__tar.gz

{ai_microcore-4.0.0.dev3 → ai_microcore-4.0.0.dev5}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: ai-microcore
-Version: 4.0.0.dev3
+Version: 4.0.0.dev5
 Summary: # Minimalistic Foundation for AI Applications
 Keywords: llm,large language models,ai,similarity search,ai search,gpt,openai
 Author-email: Vitalii Stepanenko <mail@vitalii.in>

{ai_microcore-4.0.0.dev3 → ai_microcore-4.0.0.dev5}/microcore/__init__.py

@@ -161,4 +161,4 @@ __all__ = [
     # "wrappers",
 ]
 
-__version__ = "4.0.0-dev3"
+__version__ = "4.0.0-dev5"

{ai_microcore-4.0.0.dev3 → ai_microcore-4.0.0.dev5}/microcore/_env.py

@@ -6,7 +6,7 @@ from typing import TYPE_CHECKING
 import jinja2
 
 from .embedding_db import AbstractEmbeddingDB
-from .configuration import Config, ApiType, LLMConfigError
+from .configuration import Config, ApiType, LLMConfigError, EmbeddingDbType
 from .types import TplFunctionType, LLMAsyncFunctionType, LLMFunctionType
 from .templating.jinja2 import make_jinja2_env, make_tpl_function
 from .llm.openai_llm import make_llm_functions as make_openai_llm_functions
@@ -14,6 +14,7 @@ from .llm.local_llm import make_llm_functions as make_local_llm_functions
 
 if TYPE_CHECKING:
     from .wrappers.llm_response_wrapper import LLMResponse  # noqa: F401
+    from transformers import PreTrainedModel, PreTrainedTokenizer  # noqa: F401
 
 
 @dataclass
@@ -26,10 +27,10 @@ class Env:
     llm_before_handlers: list[callable] = field(default_factory=list)
     llm_after_handlers: list[callable] = field(default_factory=list)
     texts: AbstractEmbeddingDB = None
-    model: "transformers.PreTrainedModel" = field(
+    model: "PreTrainedModel" = field(
         default=None, init=False, repr=False
     )  # noqa
-    tokenizer: "transformers.PreTrainedTokenizer" = field(  # noqa
+    tokenizer: "PreTrainedTokenizer" = field(  # noqa
         default=None, init=False, repr=False
     )
 
@@ -134,7 +135,10 @@ class Env:
             )
 
     def init_similarity_search(self):
-        if find_spec("chromadb") is not None:
+        if (
+            self.config.EMBEDDING_DB_TYPE == EmbeddingDbType.CHROMA
+            and find_spec("chromadb") is not None
+        ):
             from .embedding_db.chromadb import ChromaEmbeddingDB
 
             self.texts = ChromaEmbeddingDB(self.config)

{ai_microcore-4.0.0.dev3 → ai_microcore-4.0.0.dev5}/microcore/_llm_functions.py

@@ -1,17 +1,27 @@
+import logging
 from datetime import datetime
 
 from .utils import run_parallel
-from .wrappers.llm_response_wrapper import LLMResponse
+from .wrappers.llm_response_wrapper import LLMResponse, DictFromLLMResponse
 from .types import TPrompt
 from ._env import env
 
 
-def llm(prompt: TPrompt, **kwargs) -> str | LLMResponse:
+def llm(
+    prompt: TPrompt,
+    retries: int = 0,
+    parse_json: bool | dict = False,
+    **kwargs
+) -> str | LLMResponse:
     """
     Request Large Language Model synchronously
 
     Args:
         prompt (str | Msg | dict | list[str | Msg | dict]): Text to send to LLM
+        retries (int): Number of retries in case of error
+        parse_json (bool|dict):
+            If True, parses response as JSON,
+            alternatively non-empty dict can be used as parse_json arguments
         **kwargs (dict): Parameters supported by the LLM API
 
             See parameters supported by the OpenAI:
@@ -40,7 +50,18 @@ def llm(prompt: TPrompt, **kwargs) -> str | LLMResponse:
     """
     [h(prompt, **kwargs) for h in env().llm_before_handlers]
     start = datetime.now()
-    response = env().llm_function(prompt, **kwargs)
+    tries = retries + 1
+    while tries > 0:
+        try:
+            tries -= 1
+            response = env().llm_function(prompt, **kwargs)
+            break
+        except Exception as e:
+            if tries == 0:
+                raise e
+            logging.error(f"LLM error: {e}")
+            logging.info(f"Retrying... {tries} retries left")
+            continue
     try:
         response.gen_duration = (datetime.now() - start).total_seconds()
         if not env().config.SAVE_MEMORY:
@@ -48,15 +69,35 @@ def llm(prompt: TPrompt, **kwargs) -> str | LLMResponse:
     except AttributeError:
         ...
     [h(response) for h in env().llm_after_handlers]
+    if tries > 0:
+        retry_params = dict(**kwargs)
+        retry_params["retries"] = tries - 1
+        setattr(
+            response,
+            "_retry_callback",
+            lambda: llm(prompt, **retry_params)
+        )
+    if parse_json:
+        parsing_params = parse_json if isinstance(parse_json, dict) else {}
+        return response.parse_json(**parsing_params)
     return response
 
 
-async def allm(prompt: TPrompt, **kwargs) -> str | LLMResponse:
+async def allm(
+    prompt: TPrompt,
+    retries: int = 0,
+    parse_json: bool | dict = False,
+    **kwargs
+) -> str | LLMResponse | DictFromLLMResponse:
     """
     Request Large Language Model asynchronously
 
     Args:
         prompt (str | Msg | dict | list[str | Msg | dict]): Text to send to LLM
+        retries (int): Number of retries in case of error
+        parse_json (bool|dict):
+            If True, parses response as JSON,
+            alternatively non-empty dict can be used as parse_json arguments
         **kwargs (dict): Parameters supported by the LLM API
 
             See parameters supported by the OpenAI:
@@ -87,7 +128,18 @@ async def allm(prompt: TPrompt, **kwargs) -> str | LLMResponse:
     """
     [h(prompt, **kwargs) for h in env().llm_before_handlers]
     start = datetime.now()
-    response = await env().llm_async_function(prompt, **kwargs)
+    tries = retries + 1
+    while tries > 0:
+        try:
+            tries -= 1
+            response = await env().llm_async_function(prompt, **kwargs)
+            break
+        except Exception as e:
+            if tries == 0:
+                raise e
+            logging.error(f"LLM error: {e}")
+            logging.info(f"Retrying... {tries} retries left")
+            continue
     try:
         response.gen_duration = (datetime.now() - start).total_seconds()
         if not env().config.SAVE_MEMORY:
@@ -95,6 +147,15 @@ async def allm(prompt: TPrompt, **kwargs) -> str | LLMResponse:
     except AttributeError:
         ...
     [h(response) for h in env().llm_after_handlers]
+    if parse_json:
+        try:
+            parsing_params = parse_json if isinstance(parse_json, dict) else {}
+            return response.parse_json(**parsing_params)
+        except Exception as e:
+            if tries > 0:
+                logging.error(f"LLM error: {e}")
+                logging.info(f"Retrying... {tries} retries left")
+                return await allm(prompt, retries=tries - 1, parse_json=parse_json, **kwargs)
     return response
 
 

{ai_microcore-4.0.0.dev3 → ai_microcore-4.0.0.dev5}/microcore/ai_func/__init__.py

@@ -20,6 +20,7 @@ class AiFuncSyntax(str, Enum):
     def __str__(self):
         return self.value
 
+
 def func_arg_comments(func):
     func_source = dedent(inspect.getsource(func))
     module = ast.parse(func_source)

{ai_microcore-4.0.0.dev3 → ai_microcore-4.0.0.dev5}/microcore/configuration.py

@@ -78,6 +78,17 @@ class ApiType(str, Enum):
     def is_local(api_type: str) -> bool:
         return api_type in (ApiType.FUNCTION, ApiType.TRANSFORMERS, ApiType.NONE)
 
+    def __str__(self):
+        return self.value
+
+
+class EmbeddingDbType(str, Enum):
+    CHROMA = "chroma"
+    NONE = ""
+
+    def __str__(self):
+        return self.value
+
 
 _default_dotenv_loaded = False
 
@@ -373,6 +384,8 @@ class Config(LLMConfig):
 
     EMBEDDING_DB_PORT: str = from_env(default=None)
 
+    EMBEDDING_DB_TYPE: str = from_env(EmbeddingDbType.CHROMA)
+
     DEFAULT_ENCODING: str = from_env("utf-8")
     """Used in file system operations, utf-8 by default"""
 

{ai_microcore-4.0.0.dev3 → ai_microcore-4.0.0.dev5}/microcore/embedding_db/__init__.py

@@ -9,6 +9,7 @@ from ..utils import ExtendedString
 
 INT32_MAX = 2**31 - 1  # 2147483647
 
+
 class SearchResults(list):
     def fit_to_token_size(
         self,

{ai_microcore-4.0.0.dev3 → ai_microcore-4.0.0.dev5}/microcore/file_storage.py

@@ -15,6 +15,7 @@ from .utils import file_link, list_files
 
 _missing = object()
 
+
 @dataclass
 class Storage:
 

{ai_microcore-4.0.0.dev3 → ai_microcore-4.0.0.dev5}/microcore/json_parsing.py

@@ -102,7 +102,7 @@ def unwrap_json_substring(
         ...
 
     return (
-        input_string[start : end + 1]
+        input_string[start: end + 1]
         if brace
         else input_string if return_original_on_fail else ""
     )

{ai_microcore-4.0.0.dev3 → ai_microcore-4.0.0.dev5}/microcore/llm/_openai_llm_v0.py

@@ -9,6 +9,7 @@ from ..wrappers.llm_response_wrapper import LLMResponse
 from ..utils import is_chat_model
 from .shared import prepare_callbacks
 
+
 def _get_chunk_text(chunk, mode_chat_model: bool):
     # Azure API gives first chunk with empty choices
     choice = chunk.choices[0] if len(chunk.choices) else {}

{ai_microcore-4.0.0.dev3 → ai_microcore-4.0.0.dev5}/microcore/llm/anthropic.py

@@ -9,6 +9,7 @@ from ..types import LLMAsyncFunctionType, LLMFunctionType
 from ..wrappers.llm_response_wrapper import LLMResponse
 from .shared import prepare_callbacks
 
+
 def _get_chunk_text(chunk):
     return isinstance(chunk, ContentBlockDeltaEvent) and chunk.delta.text or ""
 
@@ -36,8 +37,15 @@ def _process_streamed_response(response, callbacks: list[callable]):
 
 
 def _prepare_llm_arguments(config: Config, kwargs: dict):
-    args = {"max_tokens": 1024, **config.LLM_DEFAULT_ARGS, **kwargs}
+    args = {**config.LLM_DEFAULT_ARGS, **kwargs}
     args["model"] = args.get("model", config.MODEL)
+    if "max_tokens" not in args:
+        if "claude-3-5-sonnet" in args["model"]:
+            args["max_tokens"] = 8192
+        elif "claude-3-7-sonnet" in args["model"]:
+            args["max_tokens"] = 16384
+        else:
+            args["max_tokens"] = 4096
     args.pop("seed", None)  # Not supported by Anthropic
     callbacks = prepare_callbacks(config, args)
     return args, {"callbacks": callbacks}

{ai_microcore-4.0.0.dev3 → ai_microcore-4.0.0.dev5}/microcore/llm/local_transformers.py

@@ -16,7 +16,7 @@ def inference(prompt: str, model, tokenizer, **kwargs):
     inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
     outputs = model.generate(**inputs, **kwargs)
     outputs = [
-        tokenizer.decode(i[len(inputs[0]) :], skip_special_tokens=skip_special_tokens)
+        tokenizer.decode(i[len(inputs[0]):], skip_special_tokens=skip_special_tokens)
         for i in outputs
     ]
     return LLMResponse(outputs[0], dict(all=outputs))

{ai_microcore-4.0.0.dev3 → ai_microcore-4.0.0.dev5}/microcore/llm/shared.py

@@ -13,6 +13,7 @@ def make_remove_hidden_output(config: Config) -> callable:
 
     return remove_hidden_output
 
+
 def prepare_callbacks(config: Config, args, set_stream: bool = True) -> list[callable]:
     callbacks = args.pop("callbacks", []) or [] + config.CALLBACKS or []
     if "callback" in args:

{ai_microcore-4.0.0.dev3 → ai_microcore-4.0.0.dev5}/microcore/text2speech/elevenlabs.py

@@ -1,9 +1,28 @@
 import os
+from dataclasses import dataclass, asdict
 from datetime import datetime
 import aiohttp
 from .._env import env
 
 
+@dataclass
+class TTSArgs:
+    text: str
+    out_file: str = None
+    voice: str = "D38z5RcWu1voky8WS1ja"
+    stability: float = 0.29
+    similarity_boost: float = 0.5
+    style: float = 0.0
+    chunk_size: int = 1024
+    speed: float = 1.0
+    use_speaker_boost: bool = False
+    previous_text: str = None
+    next_text: str = None
+
+    def to_dict(self) -> dict:
+        return asdict(self)
+
+
 async def text_to_speech(
     text: str,
     out_file: str = None,
@@ -12,6 +31,10 @@ async def text_to_speech(
     similarity_boost=0.5,
     style=0.0,
     chunk_size=1024,
+    speed=1.0,
+    use_speaker_boost: bool = False,
+    previous_text: str = None,
+    next_text: str = None,
 ) -> str:
     url = f"https://api.elevenlabs.io/v1/text-to-speech/{voice}"
     if not out_file:
@@ -25,8 +48,15 @@ async def text_to_speech(
             "stability": stability,
             "similarity_boost": similarity_boost,
             "style": style,
+            "speed": speed,
         },
     }
+    if use_speaker_boost:
+        data["voice_settings"]["use_speaker_boost"] = use_speaker_boost
+    if previous_text:
+        data["previous_text"] = previous_text
+    if next_text:
+        data["next_text"] = next_text
     headers = {
         "Accept": "audio/mpeg",
         "Content-Type": "application/json",

{ai_microcore-4.0.0.dev3 → ai_microcore-4.0.0.dev5}/microcore/tokenizing.py

@@ -5,7 +5,8 @@ import requests.exceptions
 from ._env import env
 
 
-class CantLoadTikTokenEncoding(RuntimeError): ...
+class CantLoadTikTokenEncoding(RuntimeError):
+    ...
 
 
 def _resolve_tiktoken_encoding(

{ai_microcore-4.0.0.dev3 → ai_microcore-4.0.0.dev5}/microcore/types.py

@@ -5,6 +5,7 @@ from .message_types import Msg
 
 if TYPE_CHECKING:
     from .wrappers.prompt_wrapper import PromptWrapper  # noqa: F401
+    from .wrappers.llm_response_wrapper import LLMResponse  # noqa: F401
 
 TPrompt = Union[
     dict, Msg, str, "PromptWrapper", List[Union[dict, Msg, str, "PromptWrapper"]]
@@ -12,9 +13,9 @@ TPrompt = Union[
 """Type for prompt argument in LLM requests"""
 TplFunctionType = Callable[[Union[PathLike[str], str], Any], str]
 """Function type for rendering prompt templates"""
-LLMFunctionType = Callable[[TPrompt, Any], str]
+LLMFunctionType = Callable[[TPrompt, Any], "LLMResponse"]
 """Function type for requesting LLM synchronously"""
-LLMAsyncFunctionType = Callable[[TPrompt, Any], Awaitable[str]]
+LLMAsyncFunctionType = Callable[[TPrompt, Any], Awaitable["LLMResponse"]]
 """Function type for requesting LLM asynchronously"""
 
 

{ai_microcore-4.0.0.dev3 → ai_microcore-4.0.0.dev5}/microcore/utils.py

@@ -412,9 +412,9 @@ def levenshtein(a: str, b: str) -> int:
             cost = 0 if ch_a == ch_b else 1
             current.append(
                 min(
-                    current[-1] + 1,       # insertion
-                    previous[j] + 1,       # deletion
-                    previous[j - 1] + cost # substitution
+                    current[-1] + 1,        # insertion
+                    previous[j] + 1,        # deletion
+                    previous[j - 1] + cost  # substitution
                 )
             )
         previous = current
@@ -458,4 +458,4 @@ def most_similar(
             min_dist = dist
             most_similar_word = word
 
-    return most_similar_word, min_dist
+    return most_similar_word, min_dist

{ai_microcore-4.0.0.dev3 → ai_microcore-4.0.0.dev5}/microcore/wrappers/llm_response_wrapper.py

@@ -46,9 +46,27 @@ class LLMResponse(ExtendedString, ConvertableToMessage):
         return obj
 
     def parse_json(
-        self, raise_errors: bool = True, required_fields: list[str] = None
+        self,
+        raise_errors: bool = True,
+        required_fields: list[str] = None,
+        validator: callable = None,
     ) -> list | dict | float | int | str | DictFromLLMResponse:
-        res = parse_json(self.content, raise_errors, required_fields)
+        try:
+            res = parse_json(self.content, True, required_fields)
+            if validator:
+                try:
+                    validator(res)
+                except Exception as e:
+                    raise BadAIAnswer(f"Language model response validation failed: {e}") from None
+        except Exception as e:
+            if hasattr(self, "_retry_callback"):
+                res = self._retry_callback()
+                if isinstance(res, DictFromLLMResponse):
+                    return res
+                return res.parse_json(raise_errors, required_fields, validator)
+            if raise_errors:
+                raise e
+            res = False
         if isinstance(res, dict):
             res = DictFromLLMResponse(res)
             res.llm_response = self