hamtaa-texttools 1.0.5__py3-none-any.whl → 1.1.16__py3-none-any.whl

texttools/tools/internals/operator_utils.py

@@ -0,0 +1,54 @@
+import re
+import math
+import random
+
+
+class OperatorUtils:
+    @staticmethod
+    def build_user_message(prompt: str) -> 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):
+                return []
+
+            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:
+        """
+        Calculate temperature for retry attempts.
+        """
+        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))

texttools/tools/internals/prompt_loader.py

@@ -1,4 +1,4 @@
-from typing import Optional
+from functools import lru_cache
 from pathlib import Path
 import yaml
 
@@ -7,66 +7,46 @@ class PromptLoader:
     """
     Utility for loading and formatting YAML prompt templates.
 
-    Each YAML file under `prompts/` must define at least a `main_template`,
-    and optionally an `analyze_template`. These can either be a single string
-    or a dictionary keyed by mode names (if `use_modes=True`).
-
     Responsibilities:
     - Load and parse YAML prompt definitions.
     - Select the right template (by mode, if applicable).
     - Inject variables (`{input}`, plus any extra kwargs) into the templates.
-    - Return a dict with:
-        {
-            "main_template": "...",
-            "analyze_template": "..." | None
-        }
     """
 
-    MAIN_TEMPLATE: str = "main_template"
-    ANALYZE_TEMPLATE: str = "analyze_template"
-
-    def _load_templates(
-        self,
-        prompts_dir: str,
-        prompt_file: str,
-        mode: str | None,
-    ) -> dict[str, str]:
-        prompt_path = Path(__file__).parent.parent.parent / prompts_dir / prompt_file
+    MAIN_TEMPLATE = "main_template"
+    ANALYZE_TEMPLATE = "analyze_template"
 
-        if not prompt_path.exists():
-            raise FileNotFoundError(f"Prompt file not found: {prompt_path}")
+    @staticmethod
+    def _build_format_args(text: str, **extra_kwargs) -> dict[str, str]:
+        # Base formatting args
+        format_args = {"input": text}
+        # Merge extras
+        format_args.update(extra_kwargs)
+        return format_args
 
-        try:
-            # Load the data
-            data = yaml.safe_load(prompt_path.read_text(encoding="utf-8"))
-        except yaml.YAMLError as e:
-            raise ValueError(f"Invalid YAML in {prompt_path}: {e}")
+    # Use lru_cache to load each file once
+    @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.
+        """
+        base_dir = Path(__file__).parent.parent.parent / Path("prompts")
+        prompt_path = base_dir / prompt_file
+        data = yaml.safe_load(prompt_path.read_text(encoding="utf-8"))
 
         return {
-            "main_template": data[self.MAIN_TEMPLATE][mode]
+            self.MAIN_TEMPLATE: data[self.MAIN_TEMPLATE][mode]
             if mode
             else data[self.MAIN_TEMPLATE],
-            "analyze_template": data.get(self.ANALYZE_TEMPLATE)[mode]
+            self.ANALYZE_TEMPLATE: data.get(self.ANALYZE_TEMPLATE)[mode]
             if mode
             else data.get(self.ANALYZE_TEMPLATE),
         }
 
-    def _build_format_args(self, text: str, **extra_kwargs) -> dict[str, str]:
-        # Base formatting args
-        format_args = {"input": text}
-        # Merge extras
-        format_args.update(extra_kwargs)
-        return format_args
-
     def load(
-        self,
-        prompt_file: str,
-        text: str,
-        mode: str,
-        prompts_dir: str = "prompts",
-        **extra_kwargs,
+        self, prompt_file: str, text: str, mode: str, **extra_kwargs
     ) -> dict[str, str]:
-        template_configs = self._load_templates(prompts_dir, prompt_file, mode)
+        template_configs = self._load_templates(prompt_file, mode)
         format_args = self._build_format_args(text, **extra_kwargs)
 
         # Inject variables inside each template

texttools/tools/internals/sync_operator.py

@@ -0,0 +1,201 @@
+from typing import Any, TypeVar, Type
+from collections.abc import Callable
+import logging
+
+from openai import OpenAI
+from pydantic import BaseModel
+
+from texttools.tools.internals.models import ToolOutput
+from texttools.tools.internals.operator_utils import OperatorUtils
+from texttools.tools.internals.formatters import Formatter
+from texttools.tools.internals.prompt_loader import PromptLoader
+
+# Base Model type for output models
+T = TypeVar("T", bound=BaseModel)
+
+logger = logging.getLogger("texttools.operator")
+
+
+class Operator:
+    """
+    Core engine for running text-processing operations with an LLM (Sync).
+
+    It wires together:
+    - `PromptLoader` → loads YAML prompt templates.
+    - `UserMergeFormatter` → applies formatting to messages (e.g., merging).
+    - OpenAI client → executes completions/parsed completions.
+    """
+
+    def __init__(self, client: OpenAI, model: str):
+        self._client = client
+        self._model = model
+
+    def _analyze(self, prompt_configs: dict[str, str], temperature: float) -> str:
+        """
+        Calls OpenAI API for analysis using the configured prompt template.
+        Returns the analyzed content as a string.
+        """
+        analyze_prompt = prompt_configs["analyze_template"]
+        analyze_message = [OperatorUtils.build_user_message(analyze_prompt)]
+        completion = self._client.chat.completions.create(
+            model=self._model,
+            messages=analyze_message,
+            temperature=temperature,
+        )
+        analysis = completion.choices[0].message.content.strip()
+        return analysis
+
+    def _parse_completion(
+        self,
+        message: list[dict[str, str]],
+        output_model: Type[T],
+        temperature: float,
+        logprobs: bool = False,
+        top_logprobs: int = 3,
+        priority: int | None = 0,
+    ) -> tuple[T, Any]:
+        """
+        Parses a chat completion using OpenAI's structured output format.
+        Returns both the parsed object and the raw completion for logprobs.
+        """
+        request_kwargs = {
+            "model": self._model,
+            "messages": message,
+            "response_format": output_model,
+            "temperature": temperature,
+        }
+
+        if logprobs:
+            request_kwargs["logprobs"] = True
+            request_kwargs["top_logprobs"] = top_logprobs
+
+        if priority:
+            request_kwargs["extra_body"] = {"priority": priority}
+
+        completion = self._client.beta.chat.completions.parse(**request_kwargs)
+        parsed = completion.choices[0].message.parsed
+        return parsed, completion
+
+    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 | None,
+        validator: Callable[[Any], bool] | None,
+        max_validation_retries: int | None,
+        # Internal parameters
+        prompt_file: str,
+        output_model: Type[T],
+        mode: str | None,
+        priority: int | None = 0,
+        **extra_kwargs,
+    ) -> ToolOutput:
+        """
+        Execute the LLM pipeline with the given input text.
+        """
+        prompt_loader = PromptLoader()
+        formatter = Formatter()
+        output = ToolOutput()
+        try:
+            # Prompt configs contain two keys: main_template and analyze template, both are string
+            prompt_configs = prompt_loader.load(
+                prompt_file=prompt_file,
+                text=text.strip(),
+                mode=mode,
+                **extra_kwargs,
+            )
+
+            messages = []
+
+            if with_analysis:
+                analysis = self._analyze(prompt_configs, temperature)
+                messages.append(
+                    OperatorUtils.build_user_message(
+                        f"Based on this analysis: {analysis}"
+                    )
+                )
+
+            if output_lang:
+                messages.append(
+                    OperatorUtils.build_user_message(
+                        f"Respond only in the {output_lang} language."
+                    )
+                )
+
+            if user_prompt:
+                messages.append(
+                    OperatorUtils.build_user_message(
+                        f"Consider this instruction {user_prompt}"
+                    )
+                )
+
+            messages.append(
+                OperatorUtils.build_user_message(prompt_configs["main_template"])
+            )
+
+            messages = formatter.user_merge_format(messages)
+
+            parsed, completion = self._parse_completion(
+                messages, output_model, temperature, logprobs, top_logprobs, priority
+            )
+
+            output.result = parsed.result
+
+            # Retry logic if validation fails
+            if validator and not validator(output.result):
+                for attempt in range(max_validation_retries):
+                    logger.warning(
+                        f"Validation failed, retrying for the {attempt + 1} time."
+                    )
+
+                    # Generate new temperature for retry
+                    retry_temperature = OperatorUtils.get_retry_temp(temperature)
+                    try:
+                        parsed, completion = self._parse_completion(
+                            messages,
+                            output_model,
+                            retry_temperature,
+                            logprobs,
+                            top_logprobs,
+                        )
+
+                        output.result = parsed.result
+
+                        # Check if retry was successful
+                        if validator(output.result):
+                            logger.info(
+                                f"Validation passed on retry attempt {attempt + 1}"
+                            )
+                            break
+                        else:
+                            logger.warning(
+                                f"Validation still failing after retry attempt {attempt + 1}"
+                            )
+
+                    except Exception as e:
+                        logger.error(f"Retry attempt {attempt + 1} failed: {e}")
+                        # Continue to next retry attempt if this one fails
+
+            # Final check after all retries
+            if validator and not validator(output.result):
+                output.errors.append("Validation failed after all retry attempts")
+
+            if logprobs:
+                output.logprobs = OperatorUtils.extract_logprobs(completion)
+
+            if with_analysis:
+                output.analysis = analysis
+
+            output.process = prompt_file[:-5]
+
+            return output
+
+        except Exception as e:
+            logger.error(f"TheTool failed: {e}")
+            output.errors.append(str(e))
+            return output