AI research assistant

Can an assistant that knows VBT help you write and fix research code? ChatVBT is an AI assistant for quant research in Python that brings VBT's documentation, code examples, and community discussions into the conversation. Ask how a feature works, get help adapting a backtest, or work through an error without leaving your notebook or terminal.

vbt.PF.chat("How do I add a 10% trailing stop to a signal backtest?")  

You want to add a 10% trailing stop to a signal-based portfolio backtest.

Use the `tsl_stop` argument in `vbt.PF.from_signals` and set it to `0.1`.

```python
import vectorbtpro as vbt

data = vbt.YFData.pull("BTC-USD")
fast_ma = data.run("vbt:ma", window=20, hide_params=True)
slow_ma = data.run("vbt:ma", window=50, hide_params=True)

entries = fast_ma.ma_crossed_above(slow_ma)
exits = fast_ma.ma_crossed_below(slow_ma)

pf = vbt.PF.from_signals(
    data,
    entries=entries,
    exits=exits,
    tsl_stop=0.1
)

print(pf.sharpe_ratio)
```

If you want to combine it with other stops, you can also pass
`sl_stop` and/or `tp_stop` in the same call.
...

The question went to the Portfolio class, so the model received code examples that use it as context. The returned code ran unchanged and printed a Sharpe ratio of 1.28 for the full Bitcoin history on October 2, 2026. Answers vary between runs and models, so treat them like code from a colleague: read them, run them, and check the result.

Help with the next step in your research

Start with a specific change or question. For example:

• "How do I add a trailing stop to this signal backtest?"
• "How can I test several moving-average windows across multiple assets?"
• "What does this error mean when I pass an array to my signal callback?"
• "How do I get trade statistics for each portfolio group?"

Include the relevant function, code, or error message so the answer can address your setup. You can then ask a follow-up using the same chat_history, such as adding fees or changing how assets share cash. Ask for Python code that you can run and inspect in your own research.

Choose how to work with it

Grounded chat

vbt.chat searches the whole knowledge base for the question, ranks the most relevant pages, examples, and discussions, and sends them to the model with the question. vbt.quick_chat ranks with keyword search instead of embeddings, which is faster and works offline apart from the model itself once the knowledge assets are available locally. Both take a chat_history to continue a conversation, stream the answer, and print it as plain text, render it in a notebook, or open it as an HTML page in the browser. The same commands run in a terminal as vbt chat and vbt quick-chat, as the ChatVBT highlight below shows.

Models come from OpenAI, Anthropic, Google Gemini, Hugging Face Inference, a local Ollama server, or any provider LiteLLM supports. VBT picks the first one with a key set, and every call can name a provider, a model, and the provider's own arguments. vbt.chat also needs embeddings to rank the knowledge base, which the Knowledge search page sets up.

Give the assistant your code, data, or an error message to work through a specific task. Each answer uses your question, conversation history, and relevant VBT material. Enable tools through vbt.interact to let it inspect your environment and run code. You can also use local Ollama models and local embeddings after downloading the knowledge assets. Choose the model provider separately from the search provider, so a local search setup can also be used with a hosted chat model. For changes across your own files, see Coding agents below.

Tools and reasoning

vbt.interact lets the model call tools while it answers. It can search the documentation, look up an object's API, read source code, and run Python code through the same tools as the MCP server, and you can add your own functions as tools. Supported models can show the reasoning summaries or thought segments their provider supplies. See the Function calling and Reasoning steps highlights below.

Ask any object

VBT classes with the knowledge interface can find material about themselves. find_api, find_docs, find_examples, and find_messages return its API pages, documentation, code examples, and community discussions. chat answers with the code examples as context by default, and asset_names adds the other kinds, as the Self-aware classes highlight below shows.

Coding agents

ChatVBT helps with questions and code snippets. For work across your project, coding agents such as Claude Code or Codex can work with VBT directly:

• Agent skills: vbt agents install adds step-by-step procedures for common tasks, such as generating signals, backtesting, and optimizing parameters, matched to your installed version.
• Documentation: members get llms.txt indexes, Markdown versions of every page, and downloadable knowledge assets that an agent can search.
• Tools: the MCP server, or the same functions called from Python, let an agent search the documentation and run code in your installation. See the MCP server for the available tools.
• Refactoring: vbt.refactor_source improves code with a model and VBT knowledge, as the Source refactorer highlight below shows.

✅ VBT now supports function calling with compatible models, allowing you to use the MCP server and define custom functions that the model can invoke, and all of this without any additional setup! This feature is particularly useful for complex queries, research tasks, and interactive applications where you want the model to perform specific tasks or calculations on your behalf.

env["OPENAI_API_KEY"] = "<YOUR_OPENAI_API_KEY>"

vbt.interact(  
    "How to backtest a weekly rebalancing strategy with vbt.PF.from_orders?",
    tool_display_format="compact",  
    formatter="html",
)

✅ VBT now displays reasoning summaries and thought segments from supported models, helping you follow the explanation alongside the answer. This is particularly useful for complex queries that require multiple steps to answer.

env["GITHUB_TOKEN"] = "<YOUR_GITHUB_TOKEN>"
env["OPENAI_API_KEY"] = "<YOUR_OPENAI_API_KEY>"

vbt.chat(
    "How to backtest a weekly rebalancing strategy with vbt.PF.from_orders?",
    reasoning=dict(effort="low", summary="auto"),  
    formatter="html"
)

✅ Similar to SearchVBT, ChatVBT takes search results and forwards them to your chosen model. This lets you ask questions about the VBT knowledge base and receive answers informed by the retrieved material.

env["GITHUB_TOKEN"] = "<YOUR_GITHUB_TOKEN>"
env["OPENAI_API_KEY"] = "<YOUR_OPENAI_API_KEY>"

vbt.chat("How to rebalance weekly?", formatter="html")

✅ Each VBT class offers methods to explore its features, including its API, associated documentation, Discord messages, and code examples. You can even ask it questions directly through a model!

env["GITHUB_TOKEN"] = "<YOUR_GITHUB_TOKEN>"
env["OPENAI_API_KEY"] = "<YOUR_API_KEY>"

vbt.PortfolioOptimizer.find_assets().get("link")

['https://members.vectorbt.pro/api/portfolio/pfopt/base/#vectorbtpro.portfolio.pfopt.base',
 'https://members.vectorbt.pro/api/generic/analyzable/#vectorbtpro.generic.analyzable.Analyzable',
 'https://members.vectorbt.pro/api/base/wrapping/#vectorbtpro.base.wrapping.Wrapping',
 ...
 'https://members.vectorbt.pro/features/optimization/portfolio-optimization/#riskfolio-lib',
 'https://members.vectorbt.pro/features/optimization/portfolio-optimization/#portfolio-optimization',
 'https://members.vectorbt.pro/features/optimization/portfolio-optimization/#pyportfolioopt',
 ...
 'https://discord.com/channels/x/918629995415502888/1064943203071045753',
 'https://discord.com/channels/x/918629995415502888/1067718833646874634',
 'https://discord.com/channels/x/918629995415502888/1067718855734075403',
 ...]

vbt.PortfolioOptimizer.chat("How to rebalance weekly?", formatter="html")

✅ Use the source refactorer to automatically enhance any Python code, whether it is a full package, Python object, or raw string. For larger sources, it uses the Python syntax tree to split the code into chunks, refactors each with a model, and merges the results. You can choose to return the result, update the source in place, or copy it to your clipboard. For full transparency, you can also preview the diff directly in your browser ✏️

env["GITHUB_TOKEN"] = "<YOUR_GITHUB_TOKEN>"
env["OPENAI_API_KEY"] = "<YOUR_OPENAI_API_KEY>"

source = """
@njit
def signal_func_nb(ctx, entries, exits):
    is_entry = vbt.pf_nb.select_nb(ctx, entries)
    is_exit = vbt.pf_nb.select_nb(ctx, exits)

    # TODO: Enter only if no other asset is active

    return is_entry, is_exit, False, False
"""  

new_source = vbt.refactor_source(
    source,
    attach_knowledge=True,  
    model="gpt-5-mini",
    reasoning=dict(effort="high"),
    show_diff=True,  
)
print(new_source)

@njit
def signal_func_nb(ctx, entries, exits):
    is_entry = vbt.pf_nb.select_nb(ctx, entries)
    is_exit = vbt.pf_nb.select_nb(ctx, exits)
    for col in range(ctx.from_col, ctx.to_col):
        if col != ctx.col and ctx.last_position[col] != 0:
            is_entry = False
            break
    return is_entry, is_exit, False, False