hamtaa-texttools 1.1.16__py3-none-any.whl → 1.2.0__py3-none-any.whl

texttools/core/engine.py

@@ -0,0 +1,254 @@
+import math
+import random
+import re
+from functools import lru_cache
+from pathlib import Path
+
+import yaml
+
+from .exceptions import PromptError
+
+
+class PromptLoader:
+    """
+    Utility for loading and formatting YAML prompt templates.
+    """
+
+    MAIN_TEMPLATE = "main_template"
+    ANALYZE_TEMPLATE = "analyze_template"
+
+    @lru_cache(maxsize=32)
+    def _load_templates(self, prompt_file: str, mode: str | None) -> dict[str, str]:
+        """
+        Loads prompt templates from YAML file with optional mode selection.
+        """
+        try:
+            base_dir = Path(__file__).parent.parent / Path("prompts")
+            prompt_path = base_dir / prompt_file
+
+            if not prompt_path.exists():
+                raise PromptError(f"Prompt file not found: {prompt_file}")
+
+            data = yaml.safe_load(prompt_path.read_text(encoding="utf-8"))
+
+            if self.MAIN_TEMPLATE not in data:
+                raise PromptError(f"Missing 'main_template' in {prompt_file}")
+
+            if self.ANALYZE_TEMPLATE not in data:
+                raise PromptError(f"Missing 'analyze_template' in {prompt_file}")
+
+            if mode and mode not in data.get(self.MAIN_TEMPLATE, {}):
+                raise PromptError(f"Mode '{mode}' not found in {prompt_file}")
+
+            main_template = (
+                data[self.MAIN_TEMPLATE][mode]
+                if mode and isinstance(data[self.MAIN_TEMPLATE], dict)
+                else data[self.MAIN_TEMPLATE]
+            )
+
+            analyze_template = (
+                data[self.ANALYZE_TEMPLATE][mode]
+                if mode and isinstance(data[self.ANALYZE_TEMPLATE], dict)
+                else data[self.ANALYZE_TEMPLATE]
+            )
+
+            if not main_template or not main_template.strip():
+                raise PromptError(
+                    f"Empty main_template in {prompt_file}"
+                    + (f" for mode '{mode}'" if mode else "")
+                )
+
+            return {
+                self.MAIN_TEMPLATE: main_template,
+                self.ANALYZE_TEMPLATE: analyze_template,
+            }
+
+        except yaml.YAMLError as e:
+            raise PromptError(f"Invalid YAML in {prompt_file}: {e}")
+        except Exception as e:
+            raise PromptError(f"Failed to load prompt {prompt_file}: {e}")
+
+    def load(
+        self, prompt_file: str, text: str, mode: str, **extra_kwargs
+    ) -> dict[str, str]:
+        try:
+            template_configs = self._load_templates(prompt_file, mode)
+            format_args = {"text": text}
+            format_args.update(extra_kwargs)
+
+            # Inject variables inside each template
+            for key in template_configs.keys():
+                template_configs[key] = template_configs[key].format(**format_args)
+
+            return template_configs
+
+        except KeyError as e:
+            raise PromptError(f"Missing template variable: {e}")
+        except Exception as e:
+            raise PromptError(f"Failed to format prompt: {e}")
+
+
+class OperatorUtils:
+    @staticmethod
+    def build_main_prompt(
+        main_template: str,
+        analysis: str | None,
+        output_lang: str | None,
+        user_prompt: str | None,
+    ) -> str:
+        main_prompt = ""
+
+        if analysis:
+            main_prompt += f"Based on this analysis:\n{analysis}\n"
+
+        if output_lang:
+            main_prompt += f"Respond only in the {output_lang} language.\n"
+
+        if user_prompt:
+            main_prompt += f"Consider this instruction {user_prompt}\n"
+
+        main_prompt += main_template
+
+        return main_prompt
+
+    @staticmethod
+    def build_message(prompt: str) -> list[dict[str, str]]:
+        return [{"role": "user", "content": prompt}]
+
+    @staticmethod
+    def extract_logprobs(completion: dict) -> list[dict]:
+        """
+        Extracts and filters token probabilities from completion logprobs.
+        Skips punctuation and structural tokens, returns cleaned probability data.
+        """
+        logprobs_data = []
+
+        ignore_pattern = re.compile(r'^(result|[\s\[\]\{\}",:]+)$')
+
+        for choice in completion.choices:
+            if not getattr(choice, "logprobs", None):
+                raise ValueError("Your model does not support logprobs")
+
+            for logprob_item in choice.logprobs.content:
+                if ignore_pattern.match(logprob_item.token):
+                    continue
+                token_entry = {
+                    "token": logprob_item.token,
+                    "prob": round(math.exp(logprob_item.logprob), 8),
+                    "top_alternatives": [],
+                }
+                for alt in logprob_item.top_logprobs:
+                    if ignore_pattern.match(alt.token):
+                        continue
+                    token_entry["top_alternatives"].append(
+                        {
+                            "token": alt.token,
+                            "prob": round(math.exp(alt.logprob), 8),
+                        }
+                    )
+                logprobs_data.append(token_entry)
+
+        return logprobs_data
+
+    @staticmethod
+    def get_retry_temp(base_temp: float) -> float:
+        delta_temp = random.choice([-1, 1]) * random.uniform(0.1, 0.9)
+        new_temp = base_temp + delta_temp
+
+        return max(0.0, min(new_temp, 1.5))
+
+
+def text_to_chunks(text: str, size: int, overlap: int) -> list[str]:
+    separators = ["\n\n", "\n", " ", ""]
+    is_separator_regex = False
+    keep_separator = True  # Equivalent to 'start'
+    length_function = len
+    strip_whitespace = True
+    chunk_size = size
+    chunk_overlap = overlap
+
+    def _split_text_with_regex(
+        text: str, separator: str, keep_separator: bool
+    ) -> list[str]:
+        if not separator:
+            return [text]
+        if not keep_separator:
+            return re.split(separator, text)
+        _splits = re.split(f"({separator})", text)
+        splits = [_splits[i] + _splits[i + 1] for i in range(1, len(_splits), 2)]
+        if len(_splits) % 2 == 0:
+            splits += [_splits[-1]]
+        return [_splits[0]] + splits if _splits[0] else splits
+
+    def _join_docs(docs: list[str], separator: str) -> str | None:
+        text = separator.join(docs)
+        if strip_whitespace:
+            text = text.strip()
+        return text if text else None
+
+    def _merge_splits(splits: list[str], separator: str) -> list[str]:
+        separator_len = length_function(separator)
+        docs = []
+        current_doc = []
+        total = 0
+        for d in splits:
+            len_ = length_function(d)
+            if total + len_ + (separator_len if current_doc else 0) > chunk_size:
+                if total > chunk_size:
+                    pass
+                if current_doc:
+                    doc = _join_docs(current_doc, separator)
+                    if doc is not None:
+                        docs.append(doc)
+                    while total > chunk_overlap or (
+                        total + len_ + (separator_len if current_doc else 0)
+                        > chunk_size
+                        and total > 0
+                    ):
+                        total -= length_function(current_doc[0]) + (
+                            separator_len if len(current_doc) > 1 else 0
+                        )
+                        current_doc = current_doc[1:]
+            current_doc.append(d)
+            total += len_ + (separator_len if len(current_doc) > 1 else 0)
+        doc = _join_docs(current_doc, separator)
+        if doc is not None:
+            docs.append(doc)
+        return docs
+
+    def _split_text(text: str, separators: list[str]) -> list[str]:
+        final_chunks = []
+        separator = separators[-1]
+        new_separators = []
+        for i, _s in enumerate(separators):
+            separator_ = _s if is_separator_regex else re.escape(_s)
+            if not _s:
+                separator = _s
+                break
+            if re.search(separator_, text):
+                separator = _s
+                new_separators = separators[i + 1 :]
+                break
+        separator_ = separator if is_separator_regex else re.escape(separator)
+        splits = _split_text_with_regex(text, separator_, keep_separator)
+        _separator = "" if keep_separator else separator
+        good_splits = []
+        for s in splits:
+            if length_function(s) < chunk_size:
+                good_splits.append(s)
+            else:
+                if good_splits:
+                    merged_text = _merge_splits(good_splits, _separator)
+                    final_chunks.extend(merged_text)
+                    good_splits = []
+                if not new_separators:
+                    final_chunks.append(s)
+                else:
+                    other_info = _split_text(s, new_separators)
+                    final_chunks.extend(other_info)
+        if good_splits:
+            merged_text = _merge_splits(good_splits, _separator)
+            final_chunks.extend(merged_text)
+        return final_chunks
+
+    return _split_text(text, separators)

texttools/core/exceptions.py

@@ -0,0 +1,22 @@
+class TextToolsError(Exception):
+    """Base exception for all TextTools errors."""
+
+    pass
+
+
+class PromptError(TextToolsError):
+    """Errors related to prompt loading and formatting."""
+
+    pass
+
+
+class LLMError(TextToolsError):
+    """Errors from LLM API calls."""
+
+    pass
+
+
+class ValidationError(TextToolsError):
+    """Errors from output validation."""
+
+    pass

texttools/core/internal_models.py

@@ -0,0 +1,58 @@
+from typing import Any, Literal, Type
+
+from pydantic import BaseModel, Field, create_model
+
+
+class OperatorOutput(BaseModel):
+    result: Any
+    analysis: str | None
+    logprobs: list[dict[str, Any]] | None
+
+
+class Str(BaseModel):
+    result: str = Field(..., description="The output string", example="text")
+
+
+class Bool(BaseModel):
+    result: bool = Field(
+        ..., description="Boolean indicating the output state", example=True
+    )
+
+
+class ListStr(BaseModel):
+    result: list[str] = Field(
+        ..., description="The output list of strings", example=["text_1", "text_2"]
+    )
+
+
+class ListDictStrStr(BaseModel):
+    result: list[dict[str, str]] = Field(
+        ...,
+        description="List of dictionaries containing string key-value pairs",
+        example=[{"text": "Mohammad", "type": "PER"}, {"text": "Iran", "type": "LOC"}],
+    )
+
+
+class ReasonListStr(BaseModel):
+    reason: str = Field(..., description="Thinking process that led to the output")
+    result: list[str] = Field(
+        ..., description="The output list of strings", example=["text_1", "text_2"]
+    )
+
+
+# This function is needed to create CategorizerOutput with dynamic categories
+def create_dynamic_model(allowed_values: list[str]) -> Type[BaseModel]:
+    literal_type = Literal[*allowed_values]
+
+    CategorizerOutput = create_model(
+        "CategorizerOutput",
+        reason=(
+            str,
+            Field(
+                ..., description="Explanation of why the input belongs to the category"
+            ),
+        ),
+        result=(literal_type, Field(..., description="Predicted category label")),
+    )
+
+    return CategorizerOutput

texttools/core/operators/async_operator.py

@@ -0,0 +1,194 @@
+from collections.abc import Callable
+from typing import Any, Type, TypeVar
+
+from openai import AsyncOpenAI
+from pydantic import BaseModel
+
+from ..engine import OperatorUtils, PromptLoader
+from ..exceptions import LLMError, PromptError, TextToolsError, ValidationError
+from ..internal_models import OperatorOutput
+
+# Base Model type for output models
+T = TypeVar("T", bound=BaseModel)
+
+
+class AsyncOperator:
+    """
+    Core engine for running text-processing operations with an LLM.
+    """
+
+    def __init__(self, client: AsyncOpenAI, model: str):
+        self._client = client
+        self._model = model
+
+    async def _analyze_completion(self, analyze_message: list[dict[str, str]]) -> str:
+        try:
+            completion = await self._client.chat.completions.create(
+                model=self._model,
+                messages=analyze_message,
+            )
+
+            if not completion.choices:
+                raise LLMError("No choices returned from LLM")
+
+            analysis = completion.choices[0].message.content.strip()
+
+            if not analysis:
+                raise LLMError("Empty analysis response")
+
+            return analysis
+
+        except Exception as e:
+            if isinstance(e, (PromptError, LLMError)):
+                raise
+            raise LLMError(f"Analysis failed: {e}")
+
+    async def _parse_completion(
+        self,
+        main_message: list[dict[str, str]],
+        output_model: Type[T],
+        temperature: float,
+        logprobs: bool,
+        top_logprobs: int,
+        priority: int | None,
+    ) -> tuple[T, Any]:
+        """
+        Parses a chat completion using OpenAI's structured output format.
+        Returns both the parsed Any and the raw completion for logprobs.
+        """
+        try:
+            request_kwargs = {
+                "model": self._model,
+                "messages": main_message,
+                "response_format": output_model,
+                "temperature": temperature,
+            }
+
+            if logprobs:
+                request_kwargs["logprobs"] = True
+                request_kwargs["top_logprobs"] = top_logprobs
+
+            if priority is not None:
+                request_kwargs["extra_body"] = {"priority": priority}
+
+            completion = await self._client.beta.chat.completions.parse(
+                **request_kwargs
+            )
+
+            if not completion.choices:
+                raise LLMError("No choices returned from LLM")
+
+            parsed = completion.choices[0].message.parsed
+
+            if not parsed:
+                raise LLMError("Failed to parse LLM response")
+
+            return parsed, completion
+
+        except Exception as e:
+            if isinstance(e, LLMError):
+                raise
+            raise LLMError(f"Completion failed: {e}")
+
+    async def run(
+        self,
+        # User parameters
+        text: str,
+        with_analysis: bool,
+        output_lang: str | None,
+        user_prompt: str | None,
+        temperature: float,
+        logprobs: bool,
+        top_logprobs: int,
+        validator: Callable[[Any], bool] | None,
+        max_validation_retries: int | None,
+        priority: int | None,
+        # Internal parameters
+        tool_name: str,
+        output_model: Type[T],
+        mode: str | None,
+        **extra_kwargs,
+    ) -> OperatorOutput:
+        """
+        Execute the LLM pipeline with the given input text.
+        """
+        try:
+            prompt_loader = PromptLoader()
+            prompt_configs = prompt_loader.load(
+                prompt_file=tool_name + ".yaml",
+                text=text.strip(),
+                mode=mode,
+                **extra_kwargs,
+            )
+
+            analysis: str | None = None
+
+            if with_analysis:
+                analyze_message = OperatorUtils.build_message(
+                    prompt_configs["analyze_template"]
+                )
+                analysis = await self._analyze_completion(analyze_message)
+
+            main_message = OperatorUtils.build_message(
+                OperatorUtils.build_main_prompt(
+                    prompt_configs["main_template"], analysis, output_lang, user_prompt
+                )
+            )
+
+            parsed, completion = await self._parse_completion(
+                main_message,
+                output_model,
+                temperature,
+                logprobs,
+                top_logprobs,
+                priority,
+            )
+
+            # Retry logic if validation fails
+            if validator and not validator(parsed.result):
+                if (
+                    not isinstance(max_validation_retries, int)
+                    or max_validation_retries < 1
+                ):
+                    raise ValueError("max_validation_retries should be a positive int")
+
+                succeeded = False
+                for _ in range(max_validation_retries):
+                    # Generate a new temperature to retry
+                    retry_temperature = OperatorUtils.get_retry_temp(temperature)
+
+                    try:
+                        parsed, completion = await self._parse_completion(
+                            main_message,
+                            output_model,
+                            retry_temperature,
+                            logprobs,
+                            top_logprobs,
+                            priority=priority,
+                        )
+
+                        # Check if retry was successful
+                        if validator(parsed.result):
+                            succeeded = True
+                            break
+
+                    except LLMError:
+                        pass
+
+                if not succeeded:
+                    raise ValidationError("Validation failed after all retries")
+
+            operator_output = OperatorOutput(
+                result=parsed.result,
+                analysis=analysis if with_analysis else None,
+                logprobs=OperatorUtils.extract_logprobs(completion)
+                if logprobs
+                else None,
+            )
+
+            return operator_output
+
+        except (PromptError, LLMError, ValidationError):
+            raise
+        except Exception as e:
+            raise TextToolsError(f"Unexpected error in operator: {e}")