PandasAI Skill
Chat with your data using natural language. Ask questions about DataFrames and get
insights, visualizations, and explanations powered by LLMs.
When to Use
USE when:
- Exploring an unfamiliar dataset with open-ended natural language questions
- Generating quick visualizations from descriptive prompts
- Explaining complex data transformations to stakeholders
- Building conversational data exploration interfaces (Streamlit, Jupyter, FastAPI)
- Rapid prototyping of data analysis workflows
DON'T USE when:
- Production pipelines requiring deterministic, version-controlled outputs
- Processing highly sensitive PII without anonymization (use privacy mode)
- Performance-critical paths with very large DataFrames (>100K rows)
- Simple queries that direct pandas operations would handle faster
Install
uv add pandasai # core
uv add pandasai openai # + OpenAI backend
uv add pandasai matplotlib seaborn plotly # + visualization
export OPENAI_API_KEY="sk-..."
# or ANTHROPIC_API_KEY / GOOGLE_API_KEY for other backends
Quick Start
import pandas as pd
from pandasai import SmartDataframe
from pandasai.llm import OpenAI
df = pd.read_csv("data.csv")
smart_df = SmartDataframe(df, config={"llm": OpenAI(model="gpt-4", temperature=0)})
result = smart_df.chat("What is the total revenue by region?")
smart_df.chat("Plot a bar chart of monthly sales") # saves chart to ./charts/
print(smart_df.last_code_generated) # inspect generated code
Core Capabilities
| Capability |
API |
Notes |
| Natural language query |
smart_df.chat(question) |
Returns value, DataFrame, or chart |
| Multi-table queries |
SmartDatalake([df1, df2]).chat(q) |
Assign df.name before passing |
| Inspect code |
smart_df.last_code_generated |
Verify correctness |
| Enable caching |
config={"enable_cache": True} |
Avoids repeat LLM calls |
| Privacy mode |
config={"enforce_privacy": True} |
Anonymize sensitive columns first |
| Chart saving |
config={"save_charts": True, "save_charts_path": "./charts"} |
Auto-saves matplotlib figures |
Supported LLM Backends
| Provider |
Import class |
Env var |
| OpenAI |
from pandasai.llm import OpenAI |
OPENAI_API_KEY |
| Anthropic |
from pandasai.llm import Anthropic |
ANTHROPIC_API_KEY |
| Google Gemini |
from pandasai.llm import GoogleGemini |
GOOGLE_API_KEY |
| Azure OpenAI |
from pandasai.llm import AzureOpenAI |
AZURE_OPENAI_API_KEY |
| Local (Ollama) |
from pandasai.llm import LocalLLM |
— |
Switch backends by passing a different llm= object to config. For fallback logic,
see MultiBackendAnalyzer in references/api-reference.md.
Privacy Pattern (Quick)
import hashlib
safe_df = df.copy()
for col in ["name", "email", "ssn"]:
safe_df[col] = safe_df[col].apply(lambda x: hashlib.md5(str(x).encode()).hexdigest()[:8])
smart_df = SmartDataframe(safe_df, config={"llm": llm, "enforce_privacy": True, "enable_cache": False})
See references/api-reference.md for full PrivacyAwareAnalyzer with audit logging.
Key Gotchas
- Determinism: Set
temperature=0 for reproducible results.
- Large DataFrames: Sample to ≤10K rows before wrapping to avoid context overflow.
- Cache invalidation: Cache keys are per-question string; data changes don't auto-invalidate.
- Chart display: In headless environments set matplotlib backend to
Agg before import.
- Multi-table: Assign
df.name = "table_name" before passing list to SmartDatalake.
References
- Full code examples (all 6 capabilities):
references/api-reference.md
- Integration patterns (Streamlit, FastAPI, Jupyter):
references/api-reference.md
- Best practices, caching, cost management:
references/api-reference.md
- Troubleshooting table:
references/api-reference.md
- Upstream docs: https://docs.pandas-ai.com/
- GitHub: https://github.com/gventuri/pandas-ai
Version History
- 1.0.0 (2026-01-17): Initial release — NL queries, chart generation, multi-backend, privacy modes
1---2name: pandasai3description: Conversational data analysis using natural language queries on pandas DataFrames. Use when you want to ask plain-English questions about data, generate charts, explain transformations, or build exploratory analysis interfaces — all powered by an LLM backend. Supports OpenAI, Anthropic, Google Gemini, Azure OpenAI, and local models. Handles single DataFrames (SmartDataframe) and multi-table joins (SmartDatalake).4---56# PandasAI Skill78> Chat with your data using natural language. Ask questions about DataFrames and get9> insights, visualizations, and explanations powered by LLMs.1011## When to Use1213**USE when:**14- Exploring an unfamiliar dataset with open-ended natural language questions15- Generating quick visualizations from descriptive prompts16- Explaining complex data transformations to stakeholders17- Building conversational data exploration interfaces (Streamlit, Jupyter, FastAPI)18- Rapid prototyping of data analysis workflows1920**DON'T USE when:**21- Production pipelines requiring deterministic, version-controlled outputs22- Processing highly sensitive PII without anonymization (use privacy mode)23- Performance-critical paths with very large DataFrames (>100K rows)24- Simple queries that direct pandas operations would handle faster2526## Install2728```bash29uv add pandasai # core30uv add pandasai openai # + OpenAI backend31uv add pandasai matplotlib seaborn plotly # + visualization3233export OPENAI_API_KEY="sk-..."34# or ANTHROPIC_API_KEY / GOOGLE_API_KEY for other backends35```3637## Quick Start3839```python40import pandas as pd41from pandasai import SmartDataframe42from pandasai.llm import OpenAI4344df = pd.read_csv("data.csv")45smart_df = SmartDataframe(df, config={"llm": OpenAI(model="gpt-4", temperature=0)})4647result = smart_df.chat("What is the total revenue by region?")48smart_df.chat("Plot a bar chart of monthly sales") # saves chart to ./charts/49print(smart_df.last_code_generated) # inspect generated code50```5152## Core Capabilities5354| Capability | API | Notes |55|-----------|-----|-------|56| Natural language query | `smart_df.chat(question)` | Returns value, DataFrame, or chart |57| Multi-table queries | `SmartDatalake([df1, df2]).chat(q)` | Assign `df.name` before passing |58| Inspect code | `smart_df.last_code_generated` | Verify correctness |59| Enable caching | `config={"enable_cache": True}` | Avoids repeat LLM calls |60| Privacy mode | `config={"enforce_privacy": True}` | Anonymize sensitive columns first |61| Chart saving | `config={"save_charts": True, "save_charts_path": "./charts"}` | Auto-saves matplotlib figures |6263## Supported LLM Backends6465| Provider | Import class | Env var |66|----------|-------------|---------|67| OpenAI | `from pandasai.llm import OpenAI` | `OPENAI_API_KEY` |68| Anthropic | `from pandasai.llm import Anthropic` | `ANTHROPIC_API_KEY` |69| Google Gemini | `from pandasai.llm import GoogleGemini` | `GOOGLE_API_KEY` |70| Azure OpenAI | `from pandasai.llm import AzureOpenAI` | `AZURE_OPENAI_API_KEY` |71| Local (Ollama) | `from pandasai.llm import LocalLLM` | — |7273Switch backends by passing a different `llm=` object to `config`. For fallback logic,74see `MultiBackendAnalyzer` in `references/api-reference.md`.7576## Privacy Pattern (Quick)7778```python79import hashlib8081safe_df = df.copy()82for col in ["name", "email", "ssn"]:83 safe_df[col] = safe_df[col].apply(lambda x: hashlib.md5(str(x).encode()).hexdigest()[:8])8485smart_df = SmartDataframe(safe_df, config={"llm": llm, "enforce_privacy": True, "enable_cache": False})86```8788See `references/api-reference.md` for full `PrivacyAwareAnalyzer` with audit logging.8990## Key Gotchas9192- **Determinism**: Set `temperature=0` for reproducible results.93- **Large DataFrames**: Sample to ≤10K rows before wrapping to avoid context overflow.94- **Cache invalidation**: Cache keys are per-question string; data changes don't auto-invalidate.95- **Chart display**: In headless environments set matplotlib backend to `Agg` before import.96- **Multi-table**: Assign `df.name = "table_name"` before passing list to `SmartDatalake`.9798## References99100- Full code examples (all 6 capabilities): `references/api-reference.md`101- Integration patterns (Streamlit, FastAPI, Jupyter): `references/api-reference.md`102- Best practices, caching, cost management: `references/api-reference.md`103- Troubleshooting table: `references/api-reference.md`104- Upstream docs: https://docs.pandas-ai.com/105- GitHub: https://github.com/gventuri/pandas-ai106107## Version History108109- **1.0.0** (2026-01-17): Initial release — NL queries, chart generation, multi-backend, privacy modes
vamseeachanta