📌 NOTE: This is my work repository forked from Alchemic. I interned there for three months. In the experimental folder is the main work on the second half of the internship, mainly on the email management site Instantly.

permian-ai: AI_Agents for Kapwing Online Video Editing

1️⃣ What is permian-ai

permian-ai 🦎 is a tool that assists users in video editing, it is an automatic editing agent that allows interaction through natural language. permian-ai allows you to describe in text what you want to achieve and then the agent executes it in a Kapwing web page. permian-ai implements this process by building an AI Agent.

2️⃣ How permian-ai works

There are several kinds of agents in permian-ai, and these agents have different functions. Some agents are responsible for understanding user queries and decomposing them into steps; some agents are responsible for selenium code generation based on steps to automate web page manipulation; all agents have access to vectorstores, which are vectors of definitions and corresponding scripts about video clip tasks.

Vectorstore is a chunking of textual knowledge embedded in high-dimensional vectors (the vectorstore is a knowledge base that uses similarity matching to query relevant knowledge. 💽

Faster agents only need GPT-Index vectorstore and selenium code matching, while smarter agents can generate natural language instructions and the corresponding script code. We also work on making the AI Agent learn from the html web environment to achieve real code generation without the need to give code samples manually. 🚀 But this is extremely challenging.

Our design of permian-ai works well on video editing tasks. The results show that with good enough prompt designing and prior knowledge, the AI Agent can be competent on multi-step complex tasks. More, permian-ai has good generalizability and the architecture we designed can be used for other tasks. In the future, we also hope to use it for any professional websites, such as office websites or shopping websites, etc.

This is a flowchart of the framework for the collaboration of three sub-agents 🔩. This structure is implemented in MemoryGPTScriptAgent. The Kapwing website is launched by the chromium engine and controlled by the agent-generated selenium script code.

In a single interaction, the user's input passes through the checker, recommender and executor sub-agents and outputs the control action, i.e. the automation script code. At the bottom are the tool functions, memory pool and vector store that the agent needs to use, and we show the api services that need to be used. You can choose gpt-4 and other LLM such as Anthropic and more.

3️⃣ Requirements

First install all dependencies. We use langchain 🦜🔗, llama-index 🦙 and other good stuff.

cd permian-ai
pip install -r requirements.txt

Then, fill in the config file with your api_keys and chromedriver userdata path in config/config.yaml . You need to change the KAPWING_URL, KAPWING_PROJECT_NAME to your own, and change the qdrant and pinecone api if needed(optional).

OPENAI_API_KEY: # change to your api key
QDRANT_HOST: https://...
QDRANT_API_KEY: # change to your api key
PINECONE_API_KEY: # change to your api key
PINECONE_HOST: us-east-1-aws

KAPWING_URL: https://www.kapwing.com/folder/... # your kapwing personal page url
KAPWING_PROJECT_NAME: # your kapwing project name
USER_DATA_DIR: /Users/.../user_data # your chromedriver path which logged in kapwing

You need to sign up for Kapwing, and also make sure you have chromedriver.

4️⃣ Different Agents

You can find all agents in agents/selenium_agent.py, they are

• FastAgent first use vectorstore for semantic search and recommend three relevant options, then execute the chosen one directly without using LLM.
• RecommendAgent recommends three relevant options for user to pick, according to user objective. Also vectorstore only.
• ExecuteAgent execute the selenium code.
• GPTScriptAgent is the implementation of above framework, but without memory pool.
• MemoryGPTScriptAgent is the implementation with memory 🧠.

The essential difference between these agents is actually the corresponding vectorstore data and the different parts of the prompt design. FastAgent is highly recommended, it has a time delay of five seconds or less. GPTScriptAgent and MemoryGPTScriptAgent are based on langchain tools, so they have better effect of inference chain and better visualization.

5️⃣ All Prompts

You can find all prompts in config/prompts.yaml, check and change and play with them. We didn't use Guardrail and design the format normalization ourselves. It can be improved. When you select the agent, the prompt applied to it will be used. 🖋

• FastAgent uses prompts fast_func_prompt only.
• SimpleScriptAgent uses prompts simple_script_main_agent_template_prompt, simple_script_executor_description, simple_scripts_prompt, executor_prompt and mask_prompt.
• MemoryGPTScriptAgent uses even more prompts, you can check prompts.yaml.

6️⃣ Data and Demonstrations

You can check all the demonstration data in data/ file, they are

• frameIo_dataset.csv is the dataset from frame.io video editing website.
• functions.txt has all the selenium function names, variables and demonstrations.
• instructions_scripts.txt has all the data {USER_QUERY, INSTRUCTIONS, SCRIPTS}, and masks are designed for placeholder.
• tools_document.txt is the text version of utils/tools.py.

The dataset for evaluation is data/selenium_datasets.xlsx 📚, you can use those given Queries for play. See how to play with permian-ai below.

Also, the evaluation results are stored in logs/ file.

2023-04-26 15:17:37,846 - AgentLogger - INFO - The mean accuracy is 0.92.
2023-04-26 19:19:49,634 - AgentLogger - INFO - The mean latency is 4.89.
2023-04-26 15:17:37,846 - AgentLogger - INFO - The mean top3 accuracy is 0.94.

7️⃣ Selenium Functions

We now have several selenium functions such as add_text, change_text_color, add_sound_effect, etc. We still need to build more, or let the agent learn from them. We now define selenium function with selenium-IDE. You can see all the selenium functions in utils/tools.py.

8️⃣ Vectorstores

All of them are in utils/vecs.py. We now integrate

• gpt_index_query: use llama-index.
• qdrant_query: use Qdrant.
• pcone_query: use Pinecone.
• faiss_query: use faiss.

All of them take demonstrations from data/ to build embedding vectors.

9️⃣ How to Play and Evaluate

If you want to play with permian-ai agents with your own input video editing objective, run this with FastAgent

python agents/selenium_agent.py

You can change to other agents in main function.

For evaluation with the datasets, you can run this

python tests/unit_tests.py --model_name gpt4 --temperature 0 --memory False --fast True --vectorstore_type gpt-index --use_chromedriver True

Change the arguments for your pleasure.