app/vectorstore/base.py

"""Abstract base class for LangChain-backed vector store backends.

In the LangChain experiment branch the vector store interface changes:

- ``add_documents()`` replaces ``add(chunks, vectors)`` — the store handles
  embedding internally via the injected LangChain ``Embeddings`` object.
- ``search()`` now takes a **text query** instead of a pre-computed vector —
  again the store embeds internally.

This enables a much cleaner pipeline:
  file → loader → splitter → ``add_documents()`` → done
  query → ``search(text)`` → SearchResult list → reranker → LLM
"""

from abc import ABC, abstractmethod

from langchain_core.documents import Document

from app.models.schemas import SearchResult


class VectorStore(ABC):
    """Interface for a LangChain-backed pooled vector store.

    All implementations must enforce ``tenant_id`` isolation: a search call
    for tenant A must **never** return chunks belonging to tenant B.

    The embedding step is delegated to the LangChain ``Embeddings`` object
    injected at construction time — callers never compute vectors manually.

    Example::

        from app.vectorstore.factory import get_vectorstore
        vs = get_vectorstore()
        vs.add_documents(docs)
        results = vs.search("Victorian terrace", tenant_id="t-abc", k=10)
    """

    @abstractmethod
    def add_documents(self, documents: list[Document]) -> None:
        """Embed and insert ``documents`` into the store.

        Each document must carry ``chunk_id``, ``doc_id``, and ``tenant_id``
        fields in its ``metadata`` dict.

        Args:
            documents: LangChain :class:`Document` objects with metadata.
        """
        ...

    @abstractmethod
    def search(
        self,
        query: str,
        tenant_id: str,
        k: int = 10,
        *,
        hierarchy_level: str | None = None,
        doc_id_in: frozenset[str] | None = None,
    ) -> list[SearchResult]:
        """Retrieve top-k chunks for ``tenant_id`` nearest to ``query``.

        The store embeds the query internally — no pre-computed vector needed.

        Args:
            query: Plain-text search query.
            tenant_id: Only return chunks belonging to this tenant.
            k: Maximum number of results.
            hierarchy_level: When set, only rows with this ``hierarchy_level`` metadata.
            doc_id_in: When set, only rows whose ``doc_id`` is in this set.

        Returns:
            List of :class:`~app.models.schemas.SearchResult` ordered by
            descending similarity score.
        """
        ...

    @abstractmethod
    def delete_document(self, doc_id: str) -> None:
        """Remove all chunks belonging to ``doc_id`` from the store.

        Args:
            doc_id: Document identifier whose chunks should be deleted.
        """
        ...

    @abstractmethod
    def count(self, tenant_id: str) -> int:
        """Return total chunk count for ``tenant_id``.

        Args:
            tenant_id: Tenant identifier to count.

        Returns:
            Integer chunk count.
        """
        ...

    async def search_async(
        self,
        query: str,
        tenant_id: str,
        k: int = 10,
        *,
        hierarchy_level: str | None = None,
        doc_id_in: frozenset[str] | None = None,
    ) -> list[SearchResult]:
        """Async search; default runs sync ``search`` in a thread pool."""
        from app.async_executor import run_sync_in_executor

        return await run_sync_in_executor(
            self.search,
            query,
            tenant_id,
            k=k,
            hierarchy_level=hierarchy_level,
            doc_id_in=doc_id_in,
        )

    @abstractmethod
    def count_for_doc(self, doc_id: str) -> int:
        """Return chunk count for a specific document.

        Args:
            doc_id: Document identifier to count.

        Returns:
            Integer chunk count for that document only.
        """
        ...