Curating Her Own Examples

# labs/self_learning.py
import sqlite3
from datetime import datetime, timezone

conn = sqlite3.connect(":memory:")
conn.execute("""
    CREATE TABLE IF NOT EXISTS interactions (
        id INTEGER PRIMARY KEY AUTOINCREMENT,
        user_input TEXT NOT NULL,
        response TEXT NOT NULL,
        rating INTEGER DEFAULT 0,
        created_at TEXT NOT NULL
    )
""")
conn.execute(
    "INSERT INTO interactions (user_input, response, rating, created_at)"
    " VALUES (?, ?, ?, ?)",
    ("Hello", "Oh. You again.", 4, datetime.now(timezone.utc).isoformat()),
)
conn.commit()
print(conn.execute(
    "SELECT user_input, response, rating FROM interactions").fetchall())

$ uv run python labs/self_learning.py
[('Hello', 'Oh. You again.', 4)]

The chapter 11 habits, applied to a new table: prove the schema against :memory: before any file exists, insert through ? placeholders so an apostrophe in her snark can never break the SQL, and stamp created_at in UTC so the strings sort correctly no matter what your wall clock does. The rating column is the only new idea on this table, and the entire chapter hangs off it.

def log_interaction(conn: sqlite3.Connection, user_input: str,
                    response: str, rating: int = 0) -> None:
    conn.execute(
        "INSERT INTO interactions (user_input, response, rating, created_at)"
        " VALUES (?, ?, ?, ?)",
        (user_input, response, rating, datetime.now(timezone.utc).isoformat()),
    )
    conn.commit()

def get_high_rated_examples(conn: sqlite3.Connection,
                            min_rating: int = 3, limit: int = 5) -> list[dict]:
    rows = conn.execute(
        "SELECT user_input, response FROM interactions"
        " WHERE rating >= ? ORDER BY rating DESC, id DESC LIMIT ?",
        (min_rating, limit),
    ).fetchall()
    return [{"user": r[0], "glados": r[1]} for r in rows]

log_interaction(conn, "Are you OK?", "Define OK.", rating=3)
log_interaction(conn, "What's the weather?", "I am not a window.", rating=1)
print(get_high_rated_examples(conn))

$ uv run python labs/self_learning.py
[{'user': 'Hello', 'glados': 'Oh. You again.'}, {'user': 'Are you OK?', 'glados': 'Define OK.'}]

Three rows in, two rows out: the WHERE rating >= 3 filter is the quality gate, and the 1-rated weather line died at it, which is the whole point. ORDER BY rating DESC, id DESC puts the best exchange first and breaks ties toward newer rows, so the query is deterministic. And the tuples become dicts before they leave the function, keyed user and glados to match the example format personality.json established in chapter 9. Rows convert to named data at the boundary; downstream code should never count columns.

def build_few_shot_prompt(examples: list[dict], base_prompt: str) -> str:
    if not examples:
        return base_prompt
    shots = "\n".join(
        f"User: {e['user']}\nGLaDOS: {e['glados']}" for e in examples)
    return f"{base_prompt}\n\nHere are examples of good responses:\n{shots}"

print(build_few_shot_prompt(get_high_rated_examples(conn), "You are GLaDOS."))
print("---")
print(build_few_shot_prompt([], "You are GLaDOS."))

$ uv run python labs/self_learning.py
You are GLaDOS.

Here are examples of good responses:
User: Hello
GLaDOS: Oh. You again.
User: Are you OK?
GLaDOS: Define OK.
---
You are GLaDOS.

The second print is the one to stare at. A fresh database has no rows above the bar, and without the if not examples guard the function would append a "Here are examples of good responses:" header with nothing under it: a promise to the model that the very next line breaks. The guard makes the cold start degrade to exactly the prompt you had before this chapter existed, which is what graceful fallback means: absent data produces the old behavior, not a new artifact.

# labs/self_learning.py — full file
import sqlite3
from datetime import datetime, timezone
from pathlib import Path

DB_PATH = Path("glados/data/memory.db")

def connect(db_path: Path = DB_PATH) -> sqlite3.Connection:
    db_path.parent.mkdir(parents=True, exist_ok=True)
    conn = sqlite3.connect(db_path)
    conn.execute("""CREATE TABLE IF NOT EXISTS interactions (
        id INTEGER PRIMARY KEY AUTOINCREMENT,
        user_input TEXT NOT NULL,
        response TEXT NOT NULL,
        rating INTEGER DEFAULT 0,
        created_at TEXT NOT NULL)""")
    return conn

def log_interaction(conn: sqlite3.Connection, user_input: str,
                    response: str, rating: int = 0) -> None:
    conn.execute(
        "INSERT INTO interactions (user_input, response, rating, created_at)"
        " VALUES (?, ?, ?, ?)",
        (user_input, response, rating, datetime.now(timezone.utc).isoformat()),
    )
    conn.commit()

def update_rating(conn: sqlite3.Connection, interaction_id: int,
                  rating: int) -> None:
    conn.execute("UPDATE interactions SET rating = ? WHERE id = ?",
                 (rating, interaction_id))
    conn.commit()

def get_high_rated_examples(conn: sqlite3.Connection,
                            min_rating: int = 3, limit: int = 5) -> list[dict]:
    rows = conn.execute(
        "SELECT user_input, response FROM interactions"
        " WHERE rating >= ? ORDER BY rating DESC, id DESC LIMIT ?",
        (min_rating, limit),
    ).fetchall()
    return [{"user": r[0], "glados": r[1]} for r in rows]

def build_few_shot_prompt(examples: list[dict], base_prompt: str) -> str:
    if not examples:
        return base_prompt
    shots = "\n".join(
        f"User: {e['user']}\nGLaDOS: {e['glados']}" for e in examples)
    return f"{base_prompt}\n\nHere are examples of good responses:\n{shots}"

def main() -> None:
    conn = connect()
    log_interaction(conn, "Hello", "Oh. You again.", rating=4)
    log_interaction(conn, "Are you OK?", "Define OK.", rating=3)
    log_interaction(conn, "What's the weather?", "I am not a window.", rating=1)
    examples = get_high_rated_examples(conn)
    print(f"Examples that made the cut: {len(examples)}")
    print(build_few_shot_prompt(examples, "You are GLaDOS."))
    conn.close()

if __name__ == "__main__":
    main()

$ uv run python labs/self_learning.py
Examples that made the cut: 2
You are GLaDOS.

Here are examples of good responses:
User: Hello
GLaDOS: Oh. You again.
User: Are you OK?
GLaDOS: Define OK.

connect() points at the same glados/data/memory.db that already holds her facts and conversation summaries: one file, one backup, and CREATE TABLE IF NOT EXISTS is idempotent, so adding a table to a live database is safe on every connect. update_rating is the review tool; you score a row days after it was written, keyed by id. Wiring into the voice loop takes two lines: call log_interaction after each reply with the default rating of 0, and at session start run the personality's system_prompt through build_few_shot_prompt before the first Ollama call.