angee.knowledge.retrieval

Retrieval backend strategy and the bundled lexical backend.

A :class:~angee.knowledge.models.Vault is the search namespace and the per-namespace selection point: its retrieval_class field names one of these by a key in ANGEE_KNOWLEDGE_RETRIEVAL_CLASSES, and the vault's retrieval property resolves and binds it (through :meth:~angee.knowledge.models.Vault.retrieval_for, the one public resolution seam). Mirrors agents.InferenceBackend / storage.Backend — the owning row binds the impl in __init__ and the method (:meth:RetrievalBackend.search) carries the operation arguments; a semantic plugin (pgvector/graphrag) contributes its own backend key through autoconfig without editing this addon.

Backends return an actor-scoped Page queryset/list: REBAC row scope is applied here (apply_ambient_scope), so a caller's ambient actor only ever sees pages it may read. This module stays free of the markdown text owners — search is ORM filtering over the existing title/body columns, not parsing.

RetrievalBackend

class RetrievalBackend(ImplBase)

The search strategy one vault resolves to.

Subclasses search the bound vault's pages and return an actor-scoped Page queryset/list. Abstract base: it leaves key blank and stays out of the registry; concrete leaves register a key. A future fan-out backend (lexical + semantic, fused) is a drop-in subclass — it resolves its arms through vault.retrieval_for(key) and blends their results in :meth:search.

__init__

def __init__(vault: Any) -> None

Bind this backend to the vault whose pages it searches.

search

def search(query: str, *, first: int = 20) -> Iterable[Any]

Return the actor-visible pages in this vault matching query.

LexicalRetrievalBackend

class LexicalRetrievalBackend(RetrievalBackend)

Default backend: case-insensitive substring match over title and body.

The registry default. Title and body are matched with icontains and the result is REBAC row-scoped to the ambient actor; the body match is a sequential scan (no full-text index — that is a migration the FTS/pgvector plugin owns).

search

def search(query: str, *, first: int = 20) -> Iterable[Any]

Return up to first actor-visible pages whose title or body matches query.