Files as Data Source

"Chat with your spreadsheet" is the #1 demo in GenAI — and a minefield of silent wrong answers. Four patterns, from naive to production-grade.

▶ Watch this reel

What you'll learn

Chat with CSV/Excel
DuckDB + LLM
Code-interpreter pattern
Tabular RAG

Remember this

Never let the model compute or read raw tabular data at scale — engines compute, models interpret
DuckDB turns files into queryable tables with zero import; code-interpreter covers plots/transforms; route per question
Tabular RAG keeps whole tables (with captions) as chunks — split rows are unretrievable meaning

Why naive chat-with-CSV fails

Silent sampling (context limits) · arithmetic drift · structure flattening (merged cells, totals rows).
State coverage explicitly: 'analyzed 48,211 rows'.

DuckDB + LLM

CREATE VIEW ... SELECT * FROM 'file.parquet' — zero-copy, in-process.
DESCRIBE + sampled values into the prompt; model writes SQL, engine executes (GA-18 fences).

Code-interpreter pattern

For plots/transforms/multi-file logic SQL can't express.
Same sandbox rules as AG-03; charts as linked artifacts.
Router per question: 'how many/sum' → SQL · 'show/plot/combine' → code.

Tabular RAG

Whole table + caption + section per chunk; caption-first indexing; key columns optionally mirrored to DuckDB.

Principle: models interpret, engines compute.

Code: The tri-router: one file, three engines

async def answer_over_file(question: str, file_path: str) -> Answer:
    # 0 · one-time setup: file becomes all three things
    con = duckdb.connect()
    con.execute(f"CREATE VIEW data AS SELECT * FROM '{file_path}'")
    workspace = sandbox.put(file_path)          # code-interpreter
    chunks = index_tables_with_captions(file_path)  # tabular RAG

    # 1 · cheap router classifies the question
    route = classify(question, classes=["aggregate", "transform", "semantic"])

    if route == "aggregate":                    # DuckDB
        sql = llm(SQL_PROMPT, schema=describe(con), q=question)
        rows = con.execute(f"SELECT * FROM ({sql}) _q LIMIT 500").fetchdf()
        return llm(ANSWER_PROMPT, q=question, exact_results=rows)

    if route == "transform":                    # code interpreter
        code = llm(CODE_PROMPT, q=question, files=[workspace.name])
        out = await sandbox.run(code, timeout=60)
        return Answer(text=out.stdout, artifacts=out.files)

    return rag_answer(question, store=chunks)   # semantic → tabular RAG

# Rule of the whole reel: models interpret, engines compute.