Querying Data

Video: Watch this lesson

We're saving conversations now, so the next step is reading them back. That's what the dashboard runs on. Normally I'd open a Jupyter notebook here and poke at the data first. I'd try a few queries to see what the rows look like. Our data is small and simple, so we skip that and go straight to a script.

Create db_query.py.

Connect to the same database:

from dataclasses import dataclass

from db_init import get_db_connection
from metrics import LLMCallRecord

Fetching conversations

A query returns each row as a plain tuple. You have to remember that column 4 is the model and column 6 is the prompt. That's no fun to work with. So we convert each row back into the LLMCallRecord dataclass we already use for live calls.

A helper to convert a database row into an LLMCallRecord:

def row_to_record(row):
    return LLMCallRecord(
        model=row[4],
        prompt=row[6],
        instructions=row[5],
        answer=row[2],
        prompt_tokens=row[7],
        completion_tokens=row[8],
        total_tokens=row[9],
        response_time=row[10],
        cost=row[11],
        timestamp=row[12],
    )

Now update get_conversations to use it:

def get_conversations(limit=10):
    conn = get_db_connection()
    try:
        with conn.cursor() as cur:
            cur.execute(
                """
                SELECT id, question, answer, course, model,
                       instructions, prompt,
                       prompt_tokens, completion_tokens, total_tokens,
                       response_time, cost, timestamp
                FROM conversations
                ORDER BY timestamp DESC
                LIMIT %s
                """,
                (limit,),
            )
            rows = cur.fetchall()
    finally:
        conn.close()

    return [row_to_record(row) for row in rows]

We order by timestamp to get the most recent calls. One thing to keep in mind as the table grows: there's no index on timestamp, but there is one on id. Since ids increase over time anyway, ordering by id would be faster - or you add an index on timestamp. With a handful of rows it doesn't matter, so we leave it simple for now.

Test it:

if __name__ == "__main__":
    records = get_conversations()
    for record in records:
        print(record)

Run it:

uv run python db_query.py

The output is a wall of text, not something you'd want to read all day. Still, it proves we can pull the data back out of the database. Now we put it in front of a dashboard.

← Storing Data in PostgreSQL | Streamlit Dashboard →