Chat with a CSV, but pin a known-answer guardrail so a wrong query cannot pass

The Stack

Steps

1. 1

   Load a fixed CSV and assert the aggregation matches the known answer

   This is the guardrail: a small CSV whose average-revenue-by-region you already know (EMEA=200, AMER=400, APAC=150), computed with pure stdlib so there is no model in the loop. CI asserts the aggregation equals that known answer. In real use you run df.chat() and check its reply against these same numbers, a reply that differs is a wrong query. The df.chat() call itself is fenced.

   cat > companies.csv <<'CSV'
   name,region,revenue
   Acme,EMEA,100
   Globex,EMEA,300
   Initech,AMER,200
   Umbrella,AMER,600
   Hooli,APAC,150
   CSV
   python3 - <<'PY'
   import csv, sys
   
   by_region = {}
   with open("companies.csv", newline="") as f:
       for r in csv.DictReader(f):
           by_region.setdefault(r["region"], []).append(float(r["revenue"]))
   
   avg = {reg: sum(v) / len(v) for reg, v in by_region.items()}
   golden = {"AMER": 400.0, "APAC": 150.0, "EMEA": 200.0}
   
   if avg != golden:
       print("BAD: aggregation drifted from the known answer; got " + str(avg))
       sys.exit(1)
   
   parts = ", ".join(reg + "=" + str(avg[reg]) for reg in sorted(avg))
   print("guardrail OK: avg revenue by region " + parts + " (matches the known answer); a df.chat() reply that differs is a wrong query and must be rejected")
   PY

2. 2

   Wire the guardrail around df.chat() (the model step, not checked by CI)

   pip install pandasai pandasai-litellm, configure LiteLLM with your model and key, load the same CSV with pai.read_csv, then df.chat('What is the average revenue by region?'). Compare the reply against the golden numbers above and reject on mismatch. For data you did not author, run it inside the Docker sandbox (pip install pandasai-docker), since df.chat generates and executes Python. This step calls a model and runs generated code, so it is fenced.