shira-hermes/api/services/learner.py

"""Self-learning module — post-conversation skill extraction via meta-prompt."""

from __future__ import annotations

import json
import logging
import os

from openai import AsyncOpenAI

from api.services.skills import list_skills, view_skill, save_skill

logger = logging.getLogger("shira.learner")

# Minimum thresholds to trigger learning
MIN_TOOL_CALLS = 3
MIN_TURNS = 4

EXTRACTION_PROMPT = """\
אתה מנתח שיחות ומזהה תבניות עבודה (workflows) שניתן להפוך ל-skills לשימוש חוזר.

נתח את השיחה הבאה בין עורך דין לעוזרת AI משפטית. חפש:
- רצף פעולות שחוזר על עצמו או שנראה כתבנית כללית
- workflow מורכב שעורך דין אחר יכול להשתמש בו
- תהליך שכדאי לתעד כדי שהעוזרת תבצע אותו טוב יותר בפעם הבאה

Skills קיימים (אל תיצור כפילויות):
{existing_skills}

אם מצאת workflow חדש שראוי להיות skill, החזר JSON:
{{"create": true, "name": "skill-name-in-english", "description": "תיאור בעברית", "content": "תוכן ה-SKILL.md המלא כולל frontmatter"}}

אם מצאת שיפור ל-skill קיים, החזר:
{{"update": true, "name": "existing-skill-name", "description": "תיאור מעודכן", "content": "תוכן SKILL.md מעודכן"}}

אם אין מה ללמוד, החזר:
{{"create": false}}

חשוב: החזר רק JSON תקין, בלי טקסט נוסף.
"""


def _should_learn(messages: list[dict]) -> bool:
    """Check if conversation meets learning thresholds."""
    tool_calls = sum(1 for m in messages if m.get("role") == "tool")
    turns = sum(1 for m in messages if m.get("role") in ("user", "assistant"))
    return tool_calls >= MIN_TOOL_CALLS and turns >= MIN_TURNS


def _summarize_conversation(messages: list[dict]) -> str:
    """Create a condensed version of the conversation for the extraction prompt."""
    parts = []
    for m in messages:
        role = m.get("role", "")
        content = m.get("content", "")
        if role == "system":
            continue
        if role == "tool":
            tool_id = m.get("tool_call_id", "")
            parts.append(f"[Tool Result {tool_id}]: {content[:200]}")
        elif role == "assistant":
            # Include tool calls info if present
            tool_calls = m.get("tool_calls", [])
            if tool_calls:
                for tc in tool_calls:
                    fn = tc.get("function", {})
                    parts.append(f"[Assistant calls {fn.get('name', '?')}({fn.get('arguments', '')[:100]})]")
            elif content:
                parts.append(f"[Assistant]: {content[:300]}")
        elif role == "user":
            parts.append(f"[User]: {content[:200]}")

    return "\n".join(parts)


async def extract_skills(messages: list[dict]) -> None:
    """Analyze a completed conversation and extract skills if applicable.

    This runs as a background task — errors are logged but never raised.
    """
    if not _should_learn(messages):
        logger.debug("Conversation too short for learning (skipped)")
        return

    try:
        existing = list_skills()
        existing_summary = "\n".join(
            f'- {s["name"]}: {s["description"]}' for s in existing
        ) or "אין skills קיימים."

        conversation_text = _summarize_conversation(messages)
        prompt = EXTRACTION_PROMPT.format(existing_skills=existing_summary)

        client = AsyncOpenAI(
            base_url=os.environ.get("AI_GATEWAY_URL", "http://localhost:3000").rstrip("/") + "/v1",
            api_key=os.environ.get("AI_GATEWAY_API_KEY", ""),
        )

        response = await client.chat.completions.create(
            model=os.environ.get("CLAUDE_MODEL", "sonnet"),
            messages=[
                {"role": "system", "content": prompt},
                {"role": "user", "content": f"השיחה לניתוח:\n\n{conversation_text}"},
            ],
            max_tokens=2048,
        )

        result_text = (response.choices[0].message.content or "").strip()

        # Parse JSON — handle possible markdown wrapping
        if result_text.startswith("```"):
            result_text = result_text.split("\n", 1)[-1].rsplit("```", 1)[0].strip()

        result = json.loads(result_text)

        if result.get("create"):
            name = result["name"]
            existing_content = view_skill(name)
            if existing_content:
                logger.info("Skill '%s' already exists, skipping create", name)
                return
            content = result.get("content", "")
            if not content:
                # Build content from description
                content = (
                    f"---\nname: {name}\n"
                    f"description: {result.get('description', '')}\n"
                    f"version: 1.0.0\n---\n\n"
                    f"{result.get('description', '')}\n"
                )
            save_skill(name, content)
            logger.info("Self-learning: created skill '%s'", name)

        elif result.get("update"):
            name = result["name"]
            content = result.get("content", "")
            if content:
                save_skill(name, content)
                logger.info("Self-learning: updated skill '%s'", name)

        else:
            logger.debug("Self-learning: no new skill identified")

    except json.JSONDecodeError as e:
        logger.warning("Self-learning: failed to parse extraction result: %s", e)
    except Exception as e:
        logger.error("Self-learning error: %s", e)