synth-ai 0.2.16__py3-none-any.whl → 0.2.19__py3-none-any.whl

examples/task_apps/gepa_benchmarks/common.py

@@ -0,0 +1,260 @@
+"""Shared helpers for GEPA benchmark task apps (HotpotQA, IFBench, HoVer, PUPA)."""
+
+from __future__ import annotations
+
+import json
+import os
+import re
+from typing import Any, Iterable, Mapping, Sequence
+
+import httpx
+from fastapi import HTTPException
+
+
+def _resolve_inference_url(base_url: str) -> str:
+    """Normalise a base inference URL to the chat completions endpoint."""
+
+    normalised = (base_url or "").rstrip("/")
+    if not normalised:
+        raise RuntimeError("policy.config.inference_url required")
+    if normalised.endswith("/v1/chat/completions"):
+        return normalised
+    if normalised.endswith("/chat/completions"):
+        return normalised
+    if normalised.endswith("/v1"):
+        return f"{normalised}/chat/completions"
+    return f"{normalised}/v1/chat/completions"
+
+
+_PLACEHOLDER_PATTERN = re.compile(r"\{([^{}]+)\}")
+
+
+def _substitute_placeholders(text: str, values: Mapping[str, Any]) -> str:
+    """Replace `{placeholder}` tokens in `text` with entries from `values`."""
+
+    def _replace(match: re.Match[str]) -> str:
+        key = match.group(1)
+        replacement = values.get(key)
+        return str(replacement) if replacement is not None else match.group(0)
+
+    return _PLACEHOLDER_PATTERN.sub(_replace, text)
+
+
+def render_messages(
+    policy_config: Mapping[str, Any],
+    placeholders: Mapping[str, Any],
+    default_messages: Sequence[Mapping[str, str]],
+) -> list[dict[str, str]]:
+    """Render chat messages either from policy prompt patterns or defaults."""
+
+    prompt_config = policy_config.get("prompt") if isinstance(policy_config, Mapping) else None
+    rendered: list[dict[str, str]] = []
+    if prompt_config and isinstance(prompt_config, Mapping):
+        messages = prompt_config.get("messages")
+        if isinstance(messages, Sequence):
+            for entry in messages:
+                if not isinstance(entry, Mapping):
+                    continue
+                role = str(entry.get("role") or "user")
+                pattern = entry.get("pattern") or entry.get("content") or ""
+                content = _substitute_placeholders(str(pattern), placeholders)
+                rendered.append({"role": role, "content": content})
+    if not rendered:
+        for entry in default_messages:
+            role = str(entry.get("role") or "user")
+            pattern = entry.get("pattern") or entry.get("content") or ""
+            content = _substitute_placeholders(str(pattern), placeholders)
+            rendered.append({"role": role, "content": content})
+    return rendered
+
+
+async def call_chat_completion(
+    policy_config: Mapping[str, Any],
+    placeholders: Mapping[str, Any],
+    default_messages: Sequence[Mapping[str, str]],
+    *,
+    tool_spec: Sequence[Mapping[str, Any]] | None = None,
+    tool_choice: Mapping[str, Any] | None = None,
+    timeout: float = 60.0,
+) -> tuple[str, dict[str, Any], list[dict[str, Any]]]:
+    """Invoke an OpenAI-compatible chat/completions endpoint.
+
+    Returns:
+        response_text: The assistant message text (empty string if missing).
+        raw_response: The JSON payload from the provider.
+        messages: The messages sent to the model (after placeholder substitution).
+    """
+
+    if not isinstance(policy_config, Mapping):
+        raise RuntimeError("policy.config must be a mapping for chat completion calls")
+
+    messages = render_messages(policy_config, placeholders, default_messages)
+
+    model = policy_config.get("model")
+    if not model:
+        raise RuntimeError("policy.config.model required for rollout")
+
+    temperature = policy_config.get("temperature", 0.0)
+    max_tokens = policy_config.get("max_tokens")
+    max_completion_tokens = policy_config.get("max_completion_tokens", max_tokens or 512)
+
+    inference_url = policy_config.get("inference_url") or ""
+    final_url = _resolve_inference_url(str(inference_url))
+
+    payload: dict[str, Any] = {
+        "model": model,
+        "messages": messages,
+        "temperature": temperature,
+        "max_completion_tokens": max_completion_tokens,
+    }
+    if tool_spec:
+        payload["tools"] = list(tool_spec)
+    if tool_choice:
+        payload["tool_choice"] = tool_choice
+
+    # Prefer provider-specific keys, fall back to SYNTH/OPENAI.
+    proxy_keys = {
+        "GROQ_API_KEY": os.getenv("GROQ_API_KEY"),
+        "OPENAI_API_KEY": os.getenv("OPENAI_API_KEY"),
+        "SYNTH_API_KEY": os.getenv("SYNTH_API_KEY"),
+    }
+    api_key = next((value for value in proxy_keys.values() if value), None)
+
+    headers = {"Content-Type": "application/json"}
+    if api_key:
+        headers["Authorization"] = f"Bearer {api_key}"
+
+    async with httpx.AsyncClient(timeout=httpx.Timeout(timeout)) as client:
+        response = await client.post(final_url, json=payload, headers=headers)
+
+    try:
+        data = response.json()
+    except json.JSONDecodeError as exc:  # pragma: no cover - defensive
+        raise HTTPException(
+            status_code=502,
+            detail=f"Inference provider returned invalid JSON: {response.text[:800]}",
+        ) from exc
+
+    if response.status_code >= 500:
+        raise HTTPException(
+            status_code=502,
+            detail=f"Inference provider returned an error: {data}",
+        )
+    if response.status_code >= 400:
+        raise HTTPException(
+            status_code=400,
+            detail=f"Invalid inference request: {data}",
+        )
+
+    response_text = ""
+    choices = data.get("choices") if isinstance(data, Mapping) else None
+    if isinstance(choices, Sequence) and choices:
+        message = choices[0].get("message")
+        if isinstance(message, Mapping):
+            response_text = str(message.get("content") or "")
+
+    return response_text, data, messages
+
+
+def normalise_answer(text: str) -> str:
+    """Normalise free-form text answers (HotpotQA style)."""
+
+    lowered = text.lower()
+    # Remove punctuation and articles.
+    cleaned = re.sub(r"[^a-z0-9\s]", " ", lowered)
+    cleaned = re.sub(r"\b(a|an|the)\b", " ", cleaned)
+    cleaned = re.sub(r"\s+", " ", cleaned).strip()
+    return cleaned
+
+
+_EMOJI_PATTERN = re.compile(
+    "["
+    "\U0001F600-\U0001F64F"  # emoticons
+    "\U0001F300-\U0001F5FF"  # symbols & pictographs
+    "\U0001F680-\U0001F6FF"  # transport & map symbols
+    "\U0001F1E0-\U0001F1FF"  # flags (iOS)
+    "\U00002700-\U000027BF"
+    "\U0001F900-\U0001F9FF"
+    "\U00002600-\U000026FF"
+    "\U00002B00-\U00002BFF"
+    "]",
+    flags=re.UNICODE,
+)
+
+
+def count_emojis(text: str) -> int:
+    """Return rough count of emoji characters."""
+
+    return len(_EMOJI_PATTERN.findall(text))
+
+
+def tokenize(text: str) -> list[str]:
+    """Simple whitespace/token splitter with punctuation stripping."""
+
+    cleaned = re.sub(r"[^\w\s]", " ", text.lower())
+    return [token for token in cleaned.split() if token]
+
+
+def sentence_split(text: str) -> list[str]:
+    """Split text into sentences using punctuation heuristics."""
+
+    parts = re.split(r"(?<=[.!?])\s+", text.strip())
+    return [part.strip() for part in parts if part.strip()]
+
+
+def count_numbers(text: str) -> int:
+    """Count occurrences of numeric tokens."""
+
+    return len(re.findall(r"\b\d+(?:\.\d+)?\b", text))
+
+
+def unique_word_count(tokens: Iterable[str]) -> int:
+    """Return number of unique tokens."""
+
+    return len(set(tokens))
+
+
+PRONOUNS = {
+    "i",
+    "me",
+    "you",
+    "he",
+    "him",
+    "she",
+    "her",
+    "it",
+    "we",
+    "us",
+    "they",
+    "them",
+    "my",
+    "mine",
+    "your",
+    "yours",
+    "his",
+    "hers",
+    "its",
+    "our",
+    "ours",
+    "their",
+    "theirs",
+}
+
+
+def count_pronouns(tokens: Iterable[str]) -> int:
+    """Count pronoun tokens from a predefined list."""
+
+    return sum(1 for token in tokens if token in PRONOUNS)
+
+
+__all__ = [
+    "call_chat_completion",
+    "count_emojis",
+    "count_numbers",
+    "count_pronouns",
+    "normalise_answer",
+    "render_messages",
+    "sentence_split",
+    "tokenize",
+    "unique_word_count",
+]