---

name: conversational-analysis description: "Guides PandasAI natural-language DataFrame chat, Agent workflows, LLM configuration, response objects, generated-code repair, and legacy wrapper migration." disable-model-invocation: true metadata: disco-role: operating license: NOASSERTION

Conversational Analysis

Use this sub-skill when the user wants to ask natural-language questions of tabular data through PandasAI, configure the LLM behind .chat(), inspect typed responses, debug generated code, or migrate from legacy SmartDataframe/ SmartDatalake usage.

Fast route

1. If the user only has a pandas dataframe or a CSV/Excel file, create a PandasAI DataFrame or route dataset creation to the semantic-layer sub-skill.

2. Configure a real LLM before .chat() unless this is an offline smoke:

   import pandasai as pai
   from pandasai_litellm.litellm import LiteLLM
   
   llm = LiteLLM(model="gpt-4.1-mini", api_key="...")
   pai.config.set({"llm": llm, "save_logs": True, "verbose": False, "max_retries": 3})
   

3. Use df.chat(...) for a single dataframe, pai.chat(...) for multiple DataFrames in a fresh global conversation, or Agent(...).follow_up(...) for explicit multi-turn state.

4. Inspect response.value, response.type, and response.last_code_executed before deciding that a PandasAI run succeeded.

5. If code is untrusted or the app is public-facing, route to ../sandbox-and-security/SKILL.md.

Read next

• references/api-reference.md for verified signatures, config fields, response classes, and legacy wrapper notes.
• references/workflows.md for copyable chat, multi-dataframe, follow-up, FakeLLM, and migration recipes.
• references/response-formats.md for result dictionary rules and response object handling.
• references/troubleshooting.md when .chat() fails, generated code is rejected, or response parsing errors appear.
• scripts/offline_chat_smoke.py for a deterministic no-credential smoke test.

Core choices

Situation	Use
One dataframe and no need to keep a separate explicit agent	df.chat("question")
Several dataframes in one question	pai.chat("question", df1, df2, ...)
Multi-turn conversation with explicit memory and options	agent = Agent([...], memory_size=10); agent.chat(...); agent.follow_up(...)
Need deterministic local validation	FakeLLM plus bundled offline_chat_smoke.py
User asks about SmartDataframe or SmartDatalake	Explain deprecation and migrate to pai.DataFrame, pai.chat, or Agent

Boundaries

• Route schema creation, pai.create, pai.load, SQL views, transformations, and CSV/Excel dataset modeling to ../semantic-layer/SKILL.md.
• Route @pai.skill() and global custom function registration to ../custom-skills/SKILL.md.
• Route Docker sandbox lifecycle and custom sandbox implementations to ../sandbox-and-security/SKILL.md.
• Route the pai command-line interface to ../cli-and-project-ops/SKILL.md.

Safe validation

Run this from any environment where PandasAI is installed:

python sub-skills/conversational-analysis/scripts/offline_chat_smoke.py

The script uses FakeLLM, a tiny in-memory dataframe, and generated code that calls execute_sql_query. It should not need network, provider credentials, or large data.

Common gotchas

• .chat() fails without a configured LLM, except when using FakeLLM or a supplied config object for tests.
• Generated code must call execute_sql_query and assign a valid result dictionary. Free-form pandas code without the query function is rejected by the package's validator.
• Table names in generated SQL must match the registered dataframe schema name. Hyphenated dataset paths are transformed to underscore table names.
• Printing a chart response can open an image viewer; use .save(...) in automation.