My work

[SyntaxError4Life]

v0.0.4

---

Since version 0.0.3, I've added x.ai as a provider. I tried "grok-4-fast-reasoning" and it works as well as Qwen3-4b, but their API doesn't return the reasoning, which isn't ideal for feedback.

I tried cerebras, but their API doesn't support CoT with tools and streaming. So I gave up because "qwen-3-235b-a22b-instruct-2507" (non-CoT) doesn't allow for multiple function calls simultaneously.

With Groq, I had almost nothing to do, as with deepseek.

Logikisto

[SyntaxError4Life]

0.0.6 (0.0.5)

---

Since 0.0.4, I've added HuggingFace as a provider, as well as arguments from Venice.ai (e.g., web search, on/off reasoning, etc.).

This version is called 0.0.6 because I messed up the update and had a version conflict.

For the rest, I've decided to start coding useful things to validate functionality (e.g., gradio in HF). I'm also going to start working on documentation.

Logikisto

[SyntaxError4Life]

Since the last update, a lot has happened, and I have less time to code.

I'm working quickly on the documentation and a new module, as well as some (unnecessary) features to make the framework more interesting to use.

I also have some machine learning research to do, which I plan to publish. I also have some very old Taranis prototypes in my backups, and I'm looking into how to implement their features, as they're useful.

But overall, I'm not dedicating enough time to coding this framework...

[SyntaxError4Life]

0.1.0

---

ye

I'm working on the documentation to explain how to use it in detail.

I'm also preparing the next steps, such as adding gradio for huggingface (a HF space is in development) to the framework in the next version, as well as laying the groundwork for version 0.2 (the module Agent).

[SyntaxError4Life]

0.1.1

---

Hello

It's been quite a while since I've published anything, this weekend I was a little bored at all, I spent the last 24 hours thinking (95% of the time) and coding...

So this is a version that makes it quite easy to deploy a frontend with gradio where all you have to code is just the chat function.

Here's an example I made :

import open_taranis as T
import open_taranis.web_front as W
from API_KEY import veniceai

class gradio_agent(W.chat_fn_gradio):
    def create_stream(self, messages):

        return self.request(
            self.client,
            messages=messages,
            model=self.model,
            include_venice_system_prompt=False
        )

W.gr.ChatInterface(
    fn=gradio_agent(
    client=T.clients.veniceai(api_key=veniceai),
    request=T.clients.veniceai_request,
    model="grok-41-fast",
    _system_prompt=""
).create_fn(),
    title="web front"
).launch()

(I'm making the $500 worth of DIEM that I bought from venice.ai btw)

---

Caution :

• There is a problem when it returns CoTs with , I think I know where it comes from but avoid using it if your model returns them (here grok 4.1 fast does not return them so it's good)
• Likewise, I display the model's response in the terminal that started it in order to quickly debug if nothing comes out of the API (as with CoTs
• If you don't modify create_stream a warning will be displayed in the terminal, it only does it once and it's nothing
• Finally, there is no competition at the moment or authentication with HF, I am working on it in the future to create a robust and useful system

[SyntaxError4Life]

0.1.2

---

Salut !

I've fixed the bug that caused problems with received CoT in the Gradio (web_front) display. As a result, Gradio detects them but doesn't display them, calculates the time, and that's it. The next step is to add function calls, customize the response behavior (for example, compressing memory), and integrate other features.

Regarding gradio, I realized that my way of managing the history is causing a problem because resetting the conversation doesn't do anything; I'll address this in the next update.

---

I'm also working on a CLI agent for friends and myself, but I'm the only one paying for the APIs (thanks venice.ai). So, I've added experimental support for Ollama to allow local use.

IMPORTANT: Regardless of the model, Ollama only handles one function call at a time, and 5% of the time the call is malformed.

[SyntaxError4Life]

v0.1.3

---

Now I removed the ollama module and replaced it with openai because it works better with my framework as is.

Likewise, now the reset button in the gradio (web_front) works well, when it detects that the gradio memory is reset, it does the same for the internal memory.

Finally I took inspiration from ollama and created a function (functions_to_tools) in the framework base which allows you to transform a list of python functions into their tool definition equivalent, making everything 100 times faster to do.

---

For example:

def get_user_info(name): 
"""Fetch user information from the database by name (case-insensitive)""" 
return user_database.get(name.strip().lower(), "User not found in database")

print(T.functions_to_tools([get_user_info]))

Give:

[{'type': 'function', 'function': {'name': 'get_user_info', 'description': 'Fetch user information from the database by name (case-insensitive)', 'parameters': {'type': 'object', 'properties': {'name': {'type': 'string'}}, 'required': ['name'], 'additionalProperties': False}}}]

Unlike my version which I made by hand:

[{ "type": "function", "function": { "name": "get_user_info", "description": "Access the employee database", "parameters": { "type": "object", "properties": { "name": {"type": "string", "description": "Full name of the employee to look up"} }, "required": ["name"]}}}]

[SyntaxError4Life]

v0.1.4

---

Hey

I tried several things:

1. Deploying agents on my smartphone with Ollama in Termux, but the gradio dependency was a pain, so I removed it (not necessary).

2. Creating AI agents, but handle_streaming stopped returning the tools, so I reused the code from v0.1.1 and it works.

3. Deploying a space gradio on Hugging Face (coming soon), but some things weren't compatible. Concurrency and other things should now work.

None of this worked before, but now it does...

---

Similarly, I created a grok4-heavy-like agent, and it works well, with very little code.

For v0.1.5, I'm going to try creating many code examples as well as a gradio application on HF to see the limitations and add what I need (I already know more or less what to do).

[SyntaxError4Life]

v0.1.7 (0.1.5)

---

So...

I haven't had much time to code because I'm trying to train a LLM from scratch on my eGPU and I had to reinstall Manjaro Linux on my PC.

So far I've done a few things like:

• Adding commands such as updating the framework
• Adding a TUI to manage the framework, adding the API key to env_var, and other things

I've included an example of how to create a web search agent with Brave. It's far from perfect (coded for a friend), but it already allows you to use two tools that I've integrated directly into the framework.

You have to set environment variables to use client functions or define them raw in the code (as before), the TUI can tell you which ones are defined.

[SyntaxError4Life]

v0.2.0

---

Hey everyone (I know nobody's here but I'm pretending)

Starting with v0.2.0, I'm going to change my coding style. From time to time, I'll push new things directly into the main function, but without touching the PyPI package.

For example, I'm going to add things to the advanced examples, modify the README, and so on...

These things aren't in the package, and even if I think it's necessary, I could modify the framework's source code without updating PyPI.

---

By the way, on my laptop (Manjaro Linux w/ JDE), I've set up a shortcut to launch the TUI, it looks like this:

konsole --noclose -e zsh -c "source ~/.zshrc && /home/USER/PY_ENV/bin/python /home/USER/PY_ENV/bin/taranis open; exec zsh"

If you want to replicate this....

[SyntaxError4Life]

v0.2.1

---

For this version, I tried to code a more advanced version of the Brave agent using the agent_base class.

I managed to do it fairly easily and was able to improve it.

For the future I will focus on documentation and creating even more AI agents to improve this class.

[SyntaxError4Life]

v0.2.2

---

I'm adding unofficial support for Kimi Code to create ever more powerful agents.

To use it with agent_base you must do this:

import open_taranis as T

class Custom_Agent(T.agent_base):
    def __init__(self):
        super().__init__(is_thinking_enabled=True, yield_thinking=False)

Where is_thinking_enabled will add the management of reasoning_content in the client and yield_thinking returns the CoT tokens as normal tokens

[SyntaxError4Life]

v0.2.4

---

So, so, so...

I've updated the framework a bit to be more robust for CoT and took the opportunity to do the same for the web_front.

Now you can take a pre-built agent and send it to the function generator for gradio, and bam! The agent is running in the browser.

[SyntaxError4Life]

v0.3.0

---

Wow last message on Feb 18...

So I rewrote everything from scratch to manage it all better (all that's left is to recode all the Python libraries in C++) and allow for greater compatibility. I went from 80% of the code to just 20%.

Also, I'm adding the current state of the integrated agent; here are some details:

• It is NOT a autonomous agent.
• It can only be used via the TUI (or maybe a web client).

Regarding endpoints, I've made Mistral.ai 200% compatible, whether via their endpoint, local models, or any other endpoint with their model (it almost drove me crazy).

I've added a ton of features to the TUI, have fun with them!

PS: Yes, I'm leaving the code comments in French because I'm too lazy to change them.