AIHelpMe: "AI-Enhanced Coding Assistance for Julia"

Stable Dev Build Status Coverage Aqua

AIHelpMe harnesses the power of Julia's extensive documentation and advanced AI models to provide tailored coding guidance. By integrating with PromptingTools.jl, it offers a unique, AI-assisted approach to answering your coding queries directly in Julia's environment.

[!CAUTION] This is only a prototype! We have not tuned it yet, so your mileage may vary! Always check your results from LLMs!


  • AI-Powered Assistance: Get context-aware answers to your coding questions.
  • Easy-to-Use Interface: Simple function and macro to input your questions.
  • Flexible Querying: Use different AI models for varied insights and performance versus cost trade-offs.
  • Cost-Effective: Download pre-embedded documentation to save on API calls.
  • Uniquely Tailored: Leverage the currently loaded documentation, regardless of whether it's private or public.


To install AIHelpMe, use the Julia package manager and the package name:

using Pkg


  • Julia (version 1.10 or later).
  • Internet connection for API access.
  • OpenAI API keys with available credits. See How to Obtain API Keys.
  • For optimal performance, get also Cohere API key (free for community use) and Tavily API key (free for community use).

All setup should take less than 5 minutes!

Quick Start Guide

  1. Basic Usage:

    using AIHelpMe
    aihelp("How do I implement quicksort in Julia?")
    [ Info: Done generating response. Total cost: $0.015
    AIMessage("To implement quicksort in Julia, you can use the `sort` function with the `alg=QuickSort` argument.")

    Note: As a default, we load only the Julia documentation and docstrings for standard libraries. The default model used is GPT-4 Turbo.

    You can pretty-print the answer using pprint if you return the full RAGResult (return_all=true):

    using AIHelpMe: pprint
    result = aihelp("How do I implement quicksort in Julia?", return_all=true)
    - How do I implement quicksort in Julia?
    To implement quicksort in Julia, you can use the [5,1.0]`sort`[1,1.0] function with the [1,1.0]`alg=QuickSort`[1,1.0] argument.[2,1.0]
    1. and Related Functions/Sorting Functions
    2. and Related Functions/Sorting Functions
    3. and Related Functions/Sorting Algorithms
    4. SortingAlgorithms::/
    5. AIHelpMe::/

    Note: You can see the model cheated because it can see this very documentation...

  2. aihelp Macro:

    aihelp"how to implement quicksort in Julia?"
  3. Follow-up Questions:

    aihelp!"Can you elaborate on the `sort` function?"

    Note: The ! is required for follow-up questions. aihelp! does not add new context/more information - to do that, you need to ask a new question.

  4. Pick faster models: Eg, for simple questions, GPT 3.5 might be enough, so use the alias "gpt3t":

    aihelp"Elaborate on the `sort` function and quicksort algorithm"gpt3t
    [ Info: Done generating response. Total cost: $0.002 -->
    AIMessage("The `sort` function in programming languages, including Julia.... continues for a while!
  5. Debugging: How did you come up with that answer? Check the "context" provided to the AI model (ie, the documentation snippets that were used to generate the answer):

    # Output: Pretty-printed Question + Context + Answer with color highlights

    The color highlights show you which words were NOT supported by the provided context (magenta = completely new, blue = partially new). It's a quite and intuitive way to see which function names or variables are made up versus which ones were in the context.

    You can change the kwargs of pprint to hide the annotations or potentially even show the underlying context (snippets from the documentation):

    AIHelpMe.pprint(AIHelpMe.last_result(); add_context = true, add_scores = false)

[!TIP] Your results will significantly improve if you enable re-ranking of the context to be provided to the model (eg, aihelp(..., rerank=true)) or change pipeline to update_pipeline!(:silver). It requires setting up Cohere API key but it's free for community use.

[!TIP] Do you want to safely execute the generated code? Use AICode from PromptingTools.Experimental.AgentToolsAI.. It can executed the code in a scratch module and catch errors if they happen (eg, apply directly to AIMessage response like AICode(msg)).

Noticed some weird answers? Please let us know! See the section "Help Us Improve and Debug" in the Advanced section of the docs!

How to Obtain API Keys

OpenAI API Key:

  1. Visit OpenAI's API portal.
  2. Sign up and generate an API Key.
  3. Charge some credits ($5 minimum but that will last you for a lost time).
  4. Set it as an environment variable or a local preference in PromptingTools.jl. See the instructions.

Cohere API Key:

  1. Sign up at Cohere's registration page.
  2. After registering, visit the API keys section to obtain a free, rate-limited Trial key.
  3. Set it as an environment variable or a local preference in PromptingTools.jl. See the instructions.

Tavily API Key:

  1. Sign up at Tavily.
  2. After registering, generate an API key on the Overview page. You can get a free, rate-limited Trial key.
  3. Set it as an environment variable or a local preference in PromptingTools.jl. See the instructions.


Formulating Questions:

  • Be clear and specific for the best results. Do mention the programming language (eg, Julia) and the topic (eg, "quicksort") when it's not obvious from the context.

Example Queries:

  • Simple question: aihelp"What is a DataFrame in Julia?"
  • Using a model: aihelp"best practices for error handling in Julia"gpt4t
  • Follow-up: aihelp!"Could you provide an example?"
  • Debug errors (use err REPL variable):
## define mock function to trigger method error
f(x::Int) = x^2
# we get: ERROR: MethodError: no method matching f(::Int8)

# Help is here:
aihelp"What does this error mean? \$err" # Note the $err to interpolate the stacktrace
[ Info: Done generating response. Total cost: $0.003

AIMessage("The error message "MethodError: no method matching f(::Int8)" means that there is no method defined for function `f` that accepts an argument of type `Int8`. The error message also provides the closest candidate methods that were found, which are `f(::Any, !Matched::Any)` and `f(!Matched::Int64)` in the specified file `embed_all.jl` at lines 45 and 61, respectively.")

Extending the Knowledge Base

The package by default ships with pre-processed embeddings for all Julia standard libraries, DataFrames and PromptingTools. Thanks to the amazing Julia Artifacts system, these embeddings are downloaded/cached/loaded every time the package starts.

Note: The below functions are not yet exported. Prefix them with AIHelpMe. to use them.

Building and Updating Indexes

AIHelpMe allows users to enhance its capabilities by embedding documentation from any loaded Julia module. Utilize new_index = build_index(module) to create an index for a specific module (or a vector of modules).

To update an existing index, including newly imported packages, use new index = update_index(module) or simply update_index() to include all unrecognized modules. We will add and embed only the new documentation to avoid unnecessary duplication and cost.

Managing Indexes

Once an index is built or updated, you can choose to serialize it for later use or set it as the primary index.

To use your newly created index as the main source for queries, execute load_index!(new_index). Alternatively, load a pre-existing index from a file using load_index!(file_path).

The main index for queries is held in the global variable AIHelpMe.MAIN_INDEX.

How it works

AIHelpMe leverages PromptingTools.jl to communicate with the AI models.

We apply a Retrieval Augment Generation (RAG) pattern, ie,

  • we pre-process all available documentation (and "embed it" to convert text snippets into numbers)
  • when a question is asked, we look up the most relevant documentation snippets
  • we feed the question and the documentation snippets to the AI model
  • the AI model generates the answer

This ensures that the answers are not only based on general AI knowledge but are also specifically tailored to Julia's ecosystem and best practices.


Q: Is it expensive to embed all my documentation? A: No, embedding a comprehensive set of documentation is surprisingly cost-effective. Embedding around 170 modules, including all standard libraries and more, costs approximately 8 cents and takes less than 30 seconds. To save you money, we have already embedded the Julia standard libraries and made them available for download via Artifacts. We expect that any further knowledge base extensions should be at most a few cents (see Extending the Knowledge Base).

Q: How much does it cost to ask a question? A: Each query incurs only a fraction of a cent, depending on the length and chosen model.

Q: Can I use the Cohere Trial API Key for commercial projects? A: No, a trial key is only for testing purposes. But it takes only a few clicks to switch to Production API. The cost is only $1 per 1000 searches (!!!) and has many other benefits.

Q: How accurate are the answers? A: Like any other Generative AI answers, ie, it depends and you should always double-check.

Q: Can I use it without the internet? A: Not at the moment. It might be possible in the future, as PromptingTools.jl supports local LLMs.

Q: Why do we need Cohere API Key? A: Cohere's API is used to re-rank the best matching snippets from the documentation. It's free to use in limited quantities (ie, ~thousand requests per month), which should be enough for most users. Re-ranking improves the quality and accuracy of the answers.

Q: Why do we need Tavily API Key? A: Tavily's API is used to search the best matching snippets from the documentation. It's free to use in limited quantities (ie, ~thousand requests per month), which should be enough for most users. Searching improves the quality and accuracy of the answers.

Q: Can we use Ollama (locally-hosted) models? A: Yes, see the Advanced section in the docs.

Future Directions

AIHelpMe is continuously evolving. Future updates may include:

  • GUI interface (I'm looking at you, Stipple.jl!)
  • Better tools to trace the provenance of answers (ie, where did the answer come from?).
  • Creation of a gold standard Q&A dataset for evaluation.
  • Refinement of the RAG ingestion pipeline for optimized chunk processing and deduplication.
  • Introduction of context filtering to focus on specific modules.
  • Transition to a more sophisticated multi-turn conversation design.
  • Enhancement of the marginal information provided by the RAG context.
  • Expansion of content sources beyond docstrings, potentially including documentation sites and community resources like Discourse or Slack posts.

Please note that this is merely a pre-release - we still have a long way to go...