RAG

The RAG backbone is Knowledge (AI.rag.create) and the fluent Assistant (AI.project).

Quick path

from aire import AI

assistant = (
    AI.project("docs")
    .documents("./docs")
    .model("mock:echo")
    .embedder("local:hashing")
    .vector_store("local")
)
assistant.index()
answer = assistant.ask("What is aire?")
print(answer.text, answer.citations)

Knowledge pipeline

from aire import AI
from aire.models.base import run_sync
from aire.rag.types import Document

kb = AI.rag.create(embedder=AI.models.embedder_sync("local:hashing"))
run_sync(kb.ingest([Document(text="Auth uses API keys.", metadata={"source": "sec.md"})]))
answer = run_sync(kb.ask("How does auth work?", model="mock:echo"))

Stages (all replaceable): chunker → embedder → vector store → optional rewrite → retrieve → rerank → compress → answer model.

Stores

• Default: in-process LocalVectorStore (AI.rag.vector_store("local:default"))
• Integrations (optional extras): Qdrant, Chroma, Milvus, Weaviate, pgvector, …

Remote adapters vary in how well they honor filters/ACL — prefer local for offline tests.

Rewrite / compress / ACL / incremental

Feature	Reality (0.4.0)
Rewriters	Lexical multi-query / model-backed options; some names fall back without a model
Compressors	Truncate / extractive / model; extractive is not a full LLM extractive pack
ACL	Metadata convention + split_acl_filter / post-filter helpers; identity binding from gateway auth is thin
Incremental	AI.rag.incremental(knowledge) helpers; depends on store delete_by_document support

from aire.rag.filters import split_acl_filter

store_filter, acl = split_acl_filter({"tenant": "a", "__acl__": {"roles": ["reader"]}})

Chunkers: fixed, sentence, recursive, semantic / semantic_sentence.
get_chunker("semantic") without an embedder uses sentence-boundary packing and advertises that in describe() / a UserWarning.

GraphRAG

AI.graph.create(...) builds a KnowledgeGraph (lexical extractor by default; Neo4j optional). Community detection is label-propagation style.

Remote filtering and replacement

Remote chunk identity preserves document ID, chunk index, original ID and metadata. Qdrant uses deterministic UUIDs; see the 0.4 migration guide before reopening older collections. Equality and ACL filtering happen before the final k slice. Adapters whose metadata requires client filtering rank the full candidate set first. This costs a count/query and can hit a backend's maximum query size on large collections; backend limits raise an error rather than silently dropping authorized candidates. Use native tenant-isolated collections for large multi-tenant deployments.

PostgreSQL replacement runs in one transaction. Other remote replacements enumerate and validate the complete old generation before mutation but are not distributed transactions; concurrent writers and interruption can require reconciliation. Pinecone lexical enumeration/replacement requires the serverless list/fetch API; pod indexes that lack it return a provider error. No paid Pinecone live validation has been performed for 0.4; transport contracts cover that adapter.