brasa-db-explorer

name: brasa-db-explorer description: Connect to BrasaDB's DuckDB database and execute SQL queries to explore Brazilian financial market data. Use this skill whenever the user asks to query, explore, analyze, or study data in the brasa database, asks about available datasets or tables, wants to run SQL queries against financial data, wants to create views or combine datasets for a specific application, or mentions DuckDB, BrasaDB, or SQL in the context of this project. Also trigger when the user asks questions that could be answered by querying the data — e.g., "what stocks are in IBOV?", "show me PETR4 prices", "what's the DI1 curve?", "help me create a view with these datasets", "I have datasets A and B, combine them for X".

BrasaDB Explorer

Execute SQL queries against the brasa DuckDB database to explore, analyze, and combine Brazilian financial market data.

Setup: Creating Views

Views must exist before querying. If queries return empty results or "table not found" errors, create the views first:

uv run python -c "from brasa import create_all_views; create_all_views()"

This only needs to be done once per Python process. If you open a new terminal or restart the kernel, run it again.

If a query returns empty results but no error is raised, first verify your filter values by sampling the table without filters (for example, SELECT DISTINCT symbol FROM "input.b3-cotahist-daily" LIMIT 20) before assuming views are missing.

Connecting to DuckDB

The database file lives at $BRASA_DATA_PATH/db/brasa.duckdb. Always connect with read_only=False — DuckDB raises an error if you try to mix read-only and read-write connections in the same session, and brasa itself holds a write connection internally.

If os.environ["BRASA_DATA_PATH"] raises KeyError, inform the user that the environment variable is not set and ask them to set it to the directory where brasa data is stored before proceeding.

uv run python -c "
import duckdb, os
con = duckdb.connect(os.path.join(os.environ['BRASA_DATA_PATH'], 'db', 'brasa.duckdb'), read_only=False)
result = con.sql('''
<SQL QUERY HERE>
''')
print(result)
"

For queries that return many rows or wide tables, use .df().to_string() for full output:

uv run python -c "
import duckdb, os
con = duckdb.connect(os.path.join(os.environ['BRASA_DATA_PATH'], 'db', 'brasa.duckdb'), read_only=False)
print(con.sql('''<SQL>''').df().to_string())
"

Table Naming Convention

Tables follow the pattern "layer.dataset-name" and must be double-quoted in SQL because they contain dots and hyphens:

SELECT * FROM "input.b3-cotahist-daily" LIMIT 10

The three queryable layers are:

• input — parsed raw data (e.g., daily stock prices, company info)
• staging — transformed/enriched data (e.g., returns, index compositions)
• curated — analytics-ready unified datasets (e.g., all returns, all prices)

Discovery Workflow

Use the Key Datasets Reference for common tables listed below. Run discovery queries when the user asks about datasets not listed there, or when a query returns unexpected results.

1. List all tables

SELECT table_name
FROM information_schema.tables
WHERE table_type = 'VIEW'
ORDER BY table_name

2. Inspect a table's schema

DESCRIBE "input.b3-cotahist-daily"

3. Sample data

SELECT * FROM "input.b3-cotahist-daily" LIMIT 5

4. Check row counts and date ranges

SELECT
    COUNT(*) as rows,
    MIN(refdate) as first_date,
    MAX(refdate) as last_date
FROM "input.b3-cotahist-daily"

Key Datasets Reference

These are the most commonly used datasets, with their actual column names verified against the live database. This is a curated subset — for the complete catalog of every view grouped by topic, see docs/datasets.md in the repo, or run the discovery queries above to see every available table and its current schema. Column names are case-sensitive: some input.b3-company-* tables use camelCase (e.g. issuingCompany, codeCVM) while most processed tables use snake_case.

Prices & Trading (input layer)

Indexes (input & staging layers)

Returns (staging layer)

Macro & Rates (BCB / ANBIMA)

Corporate Events

Company & Fund Data

Note: the registry views input.b3-listed-stock-etfs, input.b3-listed-fixed-income-etfs, input.b3-listed-cripto-etfs, and input.b3-listed-reits exist but may currently hold no parquet files (queries raise an IO Error: No files found). Use staging.b3-listed-funds for fund data instead.

Equities (staging layer)

Futures & Options (staging / input)

Intraday Trades (input / staging layer)

Common Query Patterns

Filter by symbol and date range

SELECT refdate, symbol, close, volume
FROM "input.b3-cotahist-daily"
WHERE symbol = 'PETR4'
  AND refdate >= '2024-01-01'
ORDER BY refdate

Aggregate across symbols

SELECT symbol, COUNT(*) as trading_days, AVG(close) as avg_price
FROM "input.b3-cotahist-daily"
WHERE refdate >= '2024-01-01'
GROUP BY symbol
ORDER BY avg_price DESC
LIMIT 20

Join datasets

SELECT
    p.refdate,
    p.symbol,
    p.close,
    p.volume,
    r.log_return
FROM "input.b3-cotahist-daily" p
JOIN "staging.b3-equities-returns" r
    ON p.refdate = r.refdate AND p.symbol = r.symbol
WHERE p.symbol = 'VALE3'
  AND p.refdate >= '2024-01-01'
ORDER BY p.refdate

Find stocks in an index

SELECT symbol, corporation_name, specification_code
FROM "staging.b3-indexes-composition"
WHERE indexes = 'IBOV'
  AND refdate = (SELECT MAX(refdate) FROM "staging.b3-indexes-composition")
ORDER BY symbol

Get a macro series (CDI, SELIC, IPCA, ...)

SELECT refdate, value
FROM "staging.bcb-sgs"
WHERE symbol = 'CDI'
  AND refdate >= '2024-01-01'
ORDER BY refdate

Get FX rates (PTAX)

SELECT refdate, bid, ask
FROM "input.bcb-currency"
WHERE currency = 'USD'
  AND refdate >= '2024-01-01'
ORDER BY refdate

Find dividends/corporate events for a stock

staging.brasa-corporate-events has no price/refdate of trading — use ex_date for timing. Filter by event_family (CASH, STOCK, SUBSCRIPTION) or a specific event_type (e.g. CASH_DIVIDEND, INTEREST_ON_CAPITAL, SPLIT, BONUS_SHARES).

SELECT ex_date, payment_date, event_type, value_cash, yield_pct
FROM "staging.brasa-corporate-events"
WHERE symbol = 'PETR4'
  AND event_family = 'CASH'
ORDER BY ex_date DESC

Creating Derived Views

When the user wants to combine datasets for a specific application, create a persistent view:

CREATE OR REPLACE VIEW "custom.my-analysis" AS
SELECT
    p.refdate,
    p.symbol,
    p.close,
    p.volume,
    r.log_return,
    ic.indexes
FROM "input.b3-cotahist-daily" p
LEFT JOIN "staging.b3-equities-returns" r
    ON p.refdate = r.refdate AND p.symbol = r.symbol
LEFT JOIN "staging.b3-indexes-composition" ic
    ON p.symbol = ic.symbol AND p.refdate = ic.refdate
WHERE p.refdate >= '2024-01-01'

View names like "custom.my-analysis" use a dot as part of the quoted name, not as a schema separator. DuckDB stores this as a quoted object name in the default schema, so no schema creation is needed.

DuckDB SQL Features

DuckDB supports advanced SQL that is useful for financial analysis:

• Window functions: LAG, LEAD, ROW_NUMBER, RANK, running averages
• CTEs: WITH clauses for readable multi-step queries
• PIVOT/UNPIVOT: reshape data between wide and long formats
• Date functions: date_trunc, date_part, date_diff, interval arithmetic
• Statistical aggregates: stddev, variance, corr, percentile_cont
• List/array functions: list_agg, array_agg for grouping values
• QUALIFY clause: filter window function results directly

Example: Rolling volatility

SELECT
    refdate,
    symbol,
    log_return,
    STDDEV(log_return) OVER (
        PARTITION BY symbol
        ORDER BY refdate
        ROWS BETWEEN 20 PRECEDING AND CURRENT ROW
    ) as vol_21d
FROM "staging.b3-equities-returns"
WHERE symbol = 'PETR4'
  AND refdate >= '2024-01-01'
ORDER BY refdate

Guidelines

• Start with discovery (list tables, describe schemas, sample rows) before writing complex queries
• Always quote table names with double quotes due to dots and hyphens
• Use LIMIT 100 or fewer when exploring unfamiliar tables. Only omit LIMIT when the user explicitly requests full output or when row count is known to be small
• Prefer CTEs over subqueries for readability
• When the user's question is ambiguous, explore the data first and show what's available before making assumptions
• For date filtering, refdate is the standard date column across most datasets
• When creating views for the user, explain what each join and filter does
• Do not execute INSERT, UPDATE, or DELETE statements against existing brasa-managed tables. Only CREATE OR REPLACE VIEW and SELECT operations are safe; if the user requests data modification, explain that brasa manages data ingestion separately