agentpool 2.2.3__py3-none-any.whl → 2.5.0__py3-none-any.whl

agentpool_toolsets/semantic_memory_toolset.py

@@ -1,536 +0,0 @@
-"""Semantic memory toolset using TypeAgent's KnowPro for knowledge processing."""
-
-from __future__ import annotations
-
-from dataclasses import dataclass
-from datetime import UTC, datetime
-import os
-import re
-import time
-from typing import TYPE_CHECKING, Any, Literal, Self
-
-from agentpool.resource_providers import ResourceProvider
-
-
-if TYPE_CHECKING:
-    from types import TracebackType
-
-    from typeagent.knowpro import answers, query, searchlang
-    from typeagent.knowpro.answer_response_schema import AnswerResponse
-    from typeagent.knowpro.conversation_base import ConversationBase
-    from typeagent.knowpro.search_query_schema import SearchQuery
-    from typeagent.podcasts import podcast
-    from typeagent.storage.memory.semrefindex import TermToSemanticRefIndex
-    import typechat
-
-    from agentpool.agents import Agent
-    from agentpool.agents.acp_agent import ACPAgent
-    from agentpool.common_types import ModelType
-    from agentpool.tools.base import Tool
-
-
-class AgentTypeChatModel:
-    """TypeChat language model backed by agentpool Agent.
-
-    Implements the typechat.TypeChatLanguageModel protocol using an Agent
-    for LLM completions instead of direct API calls.
-    """
-
-    def __init__(self, agent: Agent[Any, str] | ACPAgent) -> None:
-        """Initialize with an Agent for completions.
-
-        Args:
-            agent: The agentpool Agent to use for LLM calls
-        """
-        self.agent = agent
-
-    async def complete(
-        self, prompt: str | list[typechat.PromptSection]
-    ) -> typechat.Success[str] | typechat.Failure:
-        """Request completion from the Agent.
-
-        Args:
-            prompt: Either a string prompt or list of PromptSection dicts
-
-        Returns:
-            Success with response text or Failure with error message
-        """
-        import typechat
-
-        try:
-            # Convert prompt sections to a single string if needed
-            if isinstance(prompt, list):
-                # Combine sections into a conversation-style prompt
-                parts: list[str] = []
-                for section in prompt:
-                    role = section["role"]
-                    content = section["content"]
-                    match role:
-                        case "system":
-                            parts.append(f"[System]: {content}")
-                        case "user":
-                            parts.append(f"[User]: {content}")
-                        case "assistant":
-                            parts.append(f"[Assistant]: {content}")
-                prompt_text = "\n\n".join(parts)
-            else:
-                prompt_text = prompt
-
-            # Run the agent and get response
-            result = await self.agent.run(prompt_text)
-            return typechat.Success(result.data)
-
-        except Exception as e:  # noqa: BLE001
-            return typechat.Failure(f"Agent completion failed: {e!r}")
-
-
-@dataclass
-class ProcessingContext:
-    """Context for TypeAgent knowledge processing."""
-
-    lang_search_options: searchlang.LanguageSearchOptions
-    answer_context_options: answers.AnswerContextOptions
-    query_context: query.QueryEvalContext[podcast.PodcastMessage, TermToSemanticRefIndex]
-    query_translator: typechat.TypeChatJsonTranslator[SearchQuery]
-    answer_translator: typechat.TypeChatJsonTranslator[AnswerResponse]
-
-
-@dataclass
-class QueryResponse:
-    """Response from a knowledge query."""
-
-    success: bool
-    answer: str
-    time_ms: int
-
-
-@dataclass
-class IngestResponse:
-    """Response from ingesting content into the knowledge base."""
-
-    success: bool
-    messages_added: int
-    semantic_refs_added: int
-    error: str | None = None
-
-
-class SemanticMemoryTools(ResourceProvider):
-    """Provider for semantic memory / knowledge processing tools.
-
-    Uses TypeAgent's KnowPro for:
-    - Semantic indexing of conversations/transcripts
-    - Natural language search queries
-    - Structured answer generation
-    """
-
-    def __init__(
-        self,
-        model: ModelType = None,
-        dbname: str | None = None,
-        name: str = "semantic_memory",
-    ) -> None:
-        """Initialize semantic memory tools.
-
-        Args:
-            model: Model to use for LLM sampling (query translation, answers)
-            dbname: SQLite database path, or None for in-memory storage
-            name: Provider name
-        """
-        super().__init__(name=name)
-        self.model = model
-        self.dbname = dbname
-        self._agent: Agent[Any, str] | None = None
-        self._context: ProcessingContext | None = None
-        self._tools: list[Tool] | None = None
-
-    async def __aenter__(self) -> Self:
-        """Initialize the agent and TypeAgent context."""
-        from agentpool import Agent
-
-        # Create minimal agent for LLM sampling
-        self._agent = Agent(model=self.model, name=f"{self.name}-sampler")
-        await self._agent.__aenter__()
-        self._context = await self._make_context()  # Build TypeAgent processing context
-        return self
-
-    async def __aexit__(
-        self,
-        exc_type: type[BaseException] | None,
-        exc_val: BaseException | None,
-        exc_tb: TracebackType | None,
-    ) -> None:
-        """Cleanup agent resources."""
-        if self._agent:
-            await self._agent.__aexit__(exc_type, exc_val, exc_tb)
-            self._agent = None
-        self._context = None
-
-    async def _make_context(self) -> ProcessingContext:
-        """Create TypeAgent processing context with our Agent-backed model."""
-        from typeagent.aitools import utils
-        from typeagent.knowpro import answers, searchlang
-        from typeagent.knowpro.answer_response_schema import AnswerResponse
-        from typeagent.knowpro.convsettings import ConversationSettings
-        from typeagent.knowpro.search_query_schema import SearchQuery
-        from typeagent.podcasts import podcast
-        from typeagent.storage.utils import create_storage_provider
-
-        if self._agent is None:
-            msg = "Agent not initialized"
-            raise RuntimeError(msg)
-
-        settings = ConversationSettings()
-        # Set up storage provider (SQLite or memory)
-        settings.storage_provider = await create_storage_provider(
-            settings.message_text_index_settings,
-            settings.related_term_index_settings,
-            self.dbname,
-            podcast.PodcastMessage,
-        )
-        lang_search_options = searchlang.LanguageSearchOptions(
-            compile_options=searchlang.LanguageQueryCompileOptions(),
-            max_message_matches=25,
-        )
-        answer_context_options = answers.AnswerContextOptions(entities_top_k=50, topics_top_k=50)
-
-        query_context = await self._load_conversation_index(settings)  # Load / create conv index
-        # Create Agent-backed TypeChat model
-        model = AgentTypeChatModel(self._agent)
-        # Create translators for structured extraction
-        query_translator = utils.create_translator(model, SearchQuery)
-        answer_translator = utils.create_translator(model, AnswerResponse)
-        return ProcessingContext(
-            lang_search_options=lang_search_options,
-            answer_context_options=answer_context_options,
-            query_context=query_context,
-            query_translator=query_translator,
-            answer_translator=answer_translator,
-        )
-
-    async def _load_conversation_index(
-        self,
-        settings: Any,
-    ) -> query.QueryEvalContext[podcast.PodcastMessage, Any]:
-        """Load conversation index from database or create empty one."""
-        from typeagent.knowpro import query
-        from typeagent.podcasts import podcast
-
-        if self.dbname is None:
-            # Try loading from default test data, or create empty
-            try:
-                conversation = await podcast.Podcast.read_from_file(
-                    "testdata/Episode_53_AdrianTchaikovsky_index",
-                    settings,
-                )
-            except FileNotFoundError:
-                conversation = await podcast.Podcast.create(settings)
-        else:
-            conversation = await podcast.Podcast.create(settings)
-
-        self._conversation = conversation
-        return query.QueryEvalContext(conversation)
-
-    @property
-    def conversation(self) -> ConversationBase[Any] | None:
-        """Get the current conversation/knowledge base."""
-        return getattr(self, "_conversation", None)
-
-    async def get_tools(self) -> list[Tool]:
-        """Get available semantic memory tools."""
-        if self._tools is not None:
-            return self._tools
-
-        self._tools = [
-            self.create_tool(self.query_knowledge, read_only=True, idempotent=True),
-            self.create_tool(self.ingest_transcript),
-            self.create_tool(self.ingest_text),
-        ]
-        return self._tools
-
-    async def query_knowledge(self, question: str) -> QueryResponse:
-        """Query the knowledge base with a natural language question.
-
-        Returns an answer synthesized from indexed conversations and documents.
-
-        Args:
-            question: Natural language question to answer
-
-        Returns:
-            QueryResponse with success status, answer text, and timing
-        """
-        from typeagent.knowpro import answers, searchlang
-        import typechat
-
-        if self._context is None:
-            return QueryResponse(success=False, answer="Semantic memory not initialized", time_ms=0)
-        t0 = time.time()
-        question = question.strip()
-        if not question:
-            dt = int((time.time() - t0) * 1000)
-            return QueryResponse(success=False, answer="No question provided", time_ms=dt)
-
-        # Stage 1-3: LLM -> proto-query, compile, execute
-        result = await searchlang.search_conversation_with_language(
-            self._context.query_context.conversation,
-            self._context.query_translator,
-            question,
-            self._context.lang_search_options,
-        )
-
-        if isinstance(result, typechat.Failure):
-            dt = int((time.time() - t0) * 1000)
-            return QueryResponse(success=False, answer=result.message, time_ms=dt)
-
-        # Stage 3a-4: ordinals -> messages/semrefs, LLM -> answer
-        _, combined_answer = await answers.generate_answers(
-            self._context.answer_translator,
-            result.value,
-            self._context.query_context.conversation,
-            question,
-            options=self._context.answer_context_options,
-        )
-
-        dt = int((time.time() - t0) * 1000)
-
-        match combined_answer.type:
-            case "NoAnswer":
-                answer = combined_answer.whyNoAnswer or "No answer found"
-                return QueryResponse(success=False, answer=answer, time_ms=dt)
-            case "Answered":
-                answer = combined_answer.answer or ""
-                return QueryResponse(success=True, answer=answer, time_ms=dt)
-            case _:
-                return QueryResponse(success=False, answer="Unexpected response type", time_ms=dt)
-
-    async def ingest_transcript(
-        self,
-        file_path: str,
-        name: str | None = None,
-        fmt: Literal["auto", "txt", "vtt"] = "auto",
-    ) -> IngestResponse:
-        """Ingest a transcript file into the knowledge base.
-
-        Supports plain text (.txt) and WebVTT (.vtt) formats.
-        The content will be indexed for semantic search.
-
-        Args:
-            file_path: Path to the transcript file
-            name: Optional name for the transcript (defaults to filename)
-            fmt: File format - "auto" detects from extension, or specify "txt"/"vtt"
-
-        Returns:
-            IngestResponse with counts of added messages and semantic refs
-        """
-        if self._context is None or self.conversation is None:
-            return IngestResponse(
-                success=False,
-                messages_added=0,
-                semantic_refs_added=0,
-                error="Semantic memory not initialized",
-            )
-
-        # Detect format
-        if fmt == "auto":
-            ext = os.path.splitext(file_path)[1].lower()  # noqa: PTH122
-            fmt = "vtt" if ext == ".vtt" else "txt"
-
-        try:
-            if fmt == "vtt":
-                result = await self._ingest_vtt_file(file_path, name)
-            else:
-                result = await self._ingest_text_file(file_path, name)
-        except Exception as e:  # noqa: BLE001
-            return IngestResponse(
-                success=False,
-                messages_added=0,
-                semantic_refs_added=0,
-                error=str(e),
-            )
-        else:
-            return result
-
-    async def _ingest_vtt_file(self, file_path: str, name: str | None) -> IngestResponse:
-        """Ingest a WebVTT file."""
-        from datetime import timedelta
-
-        from typeagent.knowpro.universal_message import (
-            UNIX_EPOCH,
-            ConversationMessage,
-            ConversationMessageMeta,
-            format_timestamp_utc,
-        )
-        import webvtt  # type: ignore[import-untyped]
-
-        assert self.conversation
-        vtt = webvtt.read(file_path)
-        messages: list[ConversationMessage] = []
-
-        for caption in vtt:
-            if not caption.text.strip():
-                continue
-
-            # Parse voice tags for speaker detection
-            from typeagent.transcripts.transcript_ingest import (
-                parse_voice_tags,
-                webvtt_timestamp_to_seconds,
-            )
-
-            raw_text = getattr(caption, "raw_text", caption.text)
-            voice_segments = parse_voice_tags(raw_text)
-
-            for speaker, text in voice_segments:
-                if not text.strip():
-                    continue
-
-                offset_seconds = webvtt_timestamp_to_seconds(caption.start)
-                ts = format_timestamp_utc(UNIX_EPOCH + timedelta(seconds=offset_seconds))
-                metadata = ConversationMessageMeta(speaker=speaker, recipients=[])
-                message = ConversationMessage(text_chunks=[text], metadata=metadata, timestamp=ts)
-                messages.append(message)
-        if not messages:
-            return IngestResponse(
-                success=False,
-                messages_added=0,
-                semantic_refs_added=0,
-                error="No messages found in VTT file",
-            )
-
-        result = await self.conversation.add_messages_with_indexing(messages)
-        return IngestResponse(
-            success=True,
-            messages_added=result.messages_added,
-            semantic_refs_added=result.semrefs_added,
-        )
-
-    async def _ingest_text_file(self, file_path: str, name: str | None) -> IngestResponse:
-        """Ingest a plain text transcript file."""
-        from typeagent.knowpro.universal_message import (
-            UNIX_EPOCH,
-            ConversationMessage,
-            ConversationMessageMeta,
-            format_timestamp_utc,
-        )
-
-        with open(file_path, encoding="utf-8") as f:  # noqa: PTH123
-            lines = f.readlines()
-
-        # Parse transcript lines with speaker detection
-        speaker_pattern = re.compile(r"^\s*(?P<speaker>[A-Z][A-Z\s]*?):\s*(?P<text>.*)$")
-
-        messages: list[ConversationMessage] = []
-        current_speaker: str | None = None
-        current_chunks: list[str] = []
-        assert self.conversation
-        for line in lines:
-            stripped = line.strip()
-            if not stripped:
-                continue
-
-            match = speaker_pattern.match(stripped)
-            if match:
-                # Save previous message if exists
-                if current_chunks:
-                    meta = ConversationMessageMeta(speaker=current_speaker, recipients=[])
-                    ts = format_timestamp_utc(UNIX_EPOCH)
-                    chunks = [" ".join(current_chunks)]
-                    message = ConversationMessage(text_chunks=chunks, metadata=meta, timestamp=ts)
-                    messages.append(message)
-                current_speaker = match.group("speaker").strip()
-                current_chunks = [match.group("text").strip()]
-            elif current_chunks:
-                current_chunks.append(stripped)
-            else:
-                # No speaker detected, use None
-                current_chunks = [stripped]
-
-        # Don't forget last message
-        if current_chunks:
-            metadata = ConversationMessageMeta(speaker=current_speaker, recipients=[])
-            chunks = [" ".join(current_chunks)]
-            ts = format_timestamp_utc(UNIX_EPOCH)
-            message = ConversationMessage(text_chunks=chunks, metadata=metadata, timestamp=ts)
-            messages.append(message)
-
-        if not messages:
-            err = "No messages found in text file"
-            return IngestResponse(success=False, messages_added=0, semantic_refs_added=0, error=err)
-        result = await self.conversation.add_messages_with_indexing(messages)
-        return IngestResponse(
-            success=True,
-            messages_added=result.messages_added,
-            semantic_refs_added=result.semrefs_added,
-        )
-
-    async def ingest_text(
-        self,
-        content: str,
-        speaker: str | None = None,
-        timestamp: str | None = None,
-    ) -> IngestResponse:
-        """Ingest raw text content into the knowledge base.
-
-        Useful for adding content from memory, APIs, or other sources.
-        Optionally specify a speaker name for attribution.
-
-        Args:
-            content: The text content to ingest
-            speaker: Optional speaker/source attribution
-            timestamp: Optional ISO timestamp (defaults to now)
-
-        Returns:
-            IngestResponse with counts of added messages and semantic refs
-        """
-        from typeagent.knowpro.universal_message import (
-            ConversationMessage,
-            ConversationMessageMeta,
-            format_timestamp_utc,
-        )
-
-        if self._context is None or self.conversation is None:
-            return IngestResponse(
-                success=False,
-                messages_added=0,
-                semantic_refs_added=0,
-                error="Semantic memory not initialized",
-            )
-
-        content = content.strip()
-        if not content:
-            return IngestResponse(
-                success=False,
-                messages_added=0,
-                semantic_refs_added=0,
-                error="No content provided",
-            )
-
-        # Use provided timestamp or current time
-        if timestamp is None:
-            timestamp = format_timestamp_utc(datetime.now(UTC))
-
-        meta = ConversationMessageMeta(speaker=speaker, recipients=[])
-        message = ConversationMessage(text_chunks=[content], metadata=meta, timestamp=timestamp)
-        try:
-            result = await self.conversation.add_messages_with_indexing([message])
-            return IngestResponse(
-                success=True,
-                messages_added=result.messages_added,
-                semantic_refs_added=result.semrefs_added,
-            )
-        except Exception as e:  # noqa: BLE001
-            return IngestResponse(
-                success=False,
-                messages_added=0,
-                semantic_refs_added=0,
-                error=str(e),
-            )
-
-
-if __name__ == "__main__":
-    import anyio
-
-    async def main() -> None:
-        async with SemanticMemoryTools(model="openai:gpt-4o-mini") as tools:
-            fns = await tools.get_tools()
-            print(f"Available tools: {[t.name for t in fns]}")
-
-    anyio.run(main)