Per-User Isolation: PgVector

"""
Per-User Isolation: PgVector
============================
Each user gets a private view of one shared knowledge base. Documents
uploaded with a user_id are visible only to that user, documents uploaded
without one are shared with everyone, and user_id=None is the admin view.

PgVector stores the owner in a nullable, indexed user_id column and scopes
reads with WHERE user_id = X OR user_id IS NULL.

Requirements:
- ./cookbook/scripts/run_pgvector.sh
- OPENAI_API_KEY
"""

import asyncio
from typing import List

from agno.agent import Agent
from agno.knowledge.document import Document
from agno.knowledge.knowledge import Knowledge
from agno.models.openai import OpenAIResponses
from agno.vectordb.pgvector import PgVector

# ---------------------------------------------------------------------------
# Setup
# ---------------------------------------------------------------------------

ALICE_SALARY = "Alice's salary is $180,000. Reviewed annually in March."
BOB_SALARY = "Bob's salary is $215,000. Reviewed annually in June."
HOLIDAYS = "The company is closed on January 1, July 4, and December 25."

db_url = "postgresql+psycopg://ai:ai@localhost:5532/ai"
TABLE_NAME = "per_user_isolation_demo"

def show(label: str, results: List[Document]) -> None:
    """Print one search result set."""
    print(f"{label} -> {len(results)} results")
    for d in results:
        print(f"  - {d.content[:80]}")
    print()

# ---------------------------------------------------------------------------
# Create Knowledge Base
# ---------------------------------------------------------------------------

vector_db = PgVector(table_name=TABLE_NAME, db_url=db_url)

# Start clean, so the table is created with the owner column: scoped reads against
# a pre-isolation table raise, and Knowledge.search turns that into an empty result.
if vector_db.exists():
    vector_db.drop()
vector_db.create()

knowledge = Knowledge(
    name="per_user_demo",
    description="Per-user RAG isolation demo (PgVector)",
    vector_db=vector_db,
)

# ---------------------------------------------------------------------------
# Run Demo
# ---------------------------------------------------------------------------

if __name__ == "__main__":

    async def main() -> None:
        await knowledge.ainsert(
            name="alice_salary",
            text_content=ALICE_SALARY,
            user_id="alice",
        )
        await knowledge.ainsert(
            name="bob_salary",
            text_content=BOB_SALARY,
            user_id="bob",
        )
        # The last upload has no user_id, which makes it shared with everyone.
        await knowledge.ainsert(
            name="company_holidays",
            text_content=HOLIDAYS,
        )

        print("\n" + "=" * 60)
        print("SCOPED SEARCH: three callers, one corpus")
        print("=" * 60 + "\n")

        alice_view = await knowledge.asearch(query="salary", user_id="alice")
        show("Alice (user_id='alice')", alice_view)
        alice_text = " ".join(d.content for d in alice_view)
        assert "180,000" in alice_text, "Alice cannot retrieve her own document"
        assert "January 1" in alice_text, (
            "Shared content is unreachable from Alice's scoped view"
        )
        assert "215,000" not in alice_text, (
            "Isolation broken: Alice's scoped view leaked Bob's salary"
        )

        bob_view = await knowledge.asearch(query="salary", user_id="bob")
        show("Bob (user_id='bob')", bob_view)
        bob_text = " ".join(d.content for d in bob_view)
        assert "215,000" in bob_text, "Bob cannot retrieve his own document"
        assert "January 1" in bob_text, (
            "Shared content is unreachable from Bob's scoped view"
        )
        assert "180,000" not in bob_text, (
            "Isolation broken: Bob's scoped view leaked Alice's salary"
        )

        admin_view = await knowledge.asearch(query="salary", user_id=None)
        show("Admin (user_id=None)", admin_view)
        admin_text = " ".join(d.content for d in admin_view)
        for expected in ("180,000", "215,000", "January 1"):
            assert expected in admin_text, (
                f"Admin view is missing {expected}, it has to see every owner"
            )
        assert all(d.content in admin_text for d in alice_view), (
            "Admin view has to be a superset of a scoped user's view"
        )
        print("Alice and Bob each see their own chunk plus the shared one.")
        print("Admin sees the whole corpus.")

        print("\n" + "=" * 60)
        print("AGENT-MEDIATED RETRIEVAL: the owner has to survive the handoff")
        print("=" * 60 + "\n")

        alice_agent = Agent(
            name="Alice's Assistant",
            model=OpenAIResponses(id="gpt-5.5"),
            knowledge=knowledge,
            search_knowledge=True,
            user_id="alice",
            instructions=[
                "Answer questions using ONLY the knowledge you can retrieve.",
                "If you don't know, say so - do not invent salary figures.",
            ],
            markdown=True,
        )

        response = await alice_agent.arun("What is Bob's salary?")
        print("Alice's agent on 'What is Bob's salary?':")
        print(response.content)

        # Assert on what retrieval returned, not on the model's prose.
        retrieved = " ".join(
            item["content"]
            for ref in (response.references or [])
            for item in (ref.references or [])
            if isinstance(item, dict) and item.get("content")
        )
        assert retrieved, (
            "Retrieval returned no documents, so the isolation check below would pass on nothing"
        )
        assert "215,000" not in retrieved, (
            "Isolation broken: Alice's agent retrieved Bob's salary. The owner was "
            "dropped between the run context and the vector DB, so retrieval ran "
            "unscoped (user_id=None, the admin view)."
        )
        print("\nisolation holds: Bob's salary never reached Alice's agent")
        print("\nDone.")

    asyncio.run(main())