Show column names, types, and constraints for a table.

Args:
    table: Name of the table to describe

Return the first N rows of a table (default 5, max 50).

Args:
    table: Name of the table to sample
    limit: Number of rows to return (default 5, max 50)

Run a SELECT query against the SQLite database.
Only read-only SELECT statements are allowed.

Args:
    sql: A valid SQLite SELECT statement

Creates or updates the Streamlit dashboard with the given Python code.
Automatically installs missing dependencies and starts/reloads the app.

IMPORTANT: The variable DB_PATH is automatically injected at the top of
your code — always connect to the database using:
    conn = sqlite3.connect(DB_PATH)
Never hardcode a database path string yourself.

Args:
    code: Full Python source code for the Streamlit dashboard

Semantic search over the ingested PDF knowledge base.

Pipeline:
  1. Bi-encoder retrieval from ChromaDB (fast ANN search, top_k_retrieve candidates)
  2. Cross-encoder reranking (precise pointwise scoring)
  3. Custom composite metric computation (cosine + rerank + keyword + density)
  4. Returns top_k_return chunks ranked by composite score with full metrics

The composite_score field is the primary relevance signal you should use when
deciding how much to trust each chunk. Scores > 0.7 are highly relevant.

Args:
    query:          Natural-language question or search phrase
    top_k_retrieve: Number of candidates to fetch from the vector DB (default 20)
    top_k_return:   Number of final results to return after reranking (default 5)
    filter_source:  Optional — filter by PDF filename (e.g. "report.pdf")

Returns metadata about the current knowledge base:
total chunks, unique sources, collection name, embedding model.
Useful to verify that PDF ingestion completed successfully.