Tool Augmented AI
Function/Tool Calling
Augmenting Large Language Models (LLMs) with Function Calls is a crucial technique that enables the integration of diverse data sources and enhances their performance. This process involves converting unstructured or noisy input data, such as natural language text, web pages, or even sensor readings into structured formats like JSON.
How to test Function/Tool Calling?
To test your JSON schema and conversations, please visit the webpage Function Calling Tool
Fine Tuning a Model for Function Calling
To train a model to have Tool Calling capability, it's a best practice to cover the conversation examples with below taxonomies/domains.
1. Turn
single turn: The user asks only one question and leave the conversation.multiple turns: The user asks follow-up questions, there's back and forth between the user and the assistant.
2. Function Call
single call: The user asks a question which requires only one function call in a row.multiple parallel calls: The user asks a question which requires multiple function calls in a row.multiple consecutive calls: The user asks a question which requires one/multiple function calls consecutively in a row.no call: The user asks a question which does not require a function call at all.
3. Adversarial User Cases
- Missing information
- Happy path vs Unhappy path
- User cancels request
- User changes params
- User asks params
- User rejects providing params
- Recover from failure
- System pivoting
- Toxic user
- Noisy input
- etc
4. Major Domains:
- Data Extraction: Noisy, Natural Language to JSON. Where the data is hidden in a huge unstructured data.
- Web Scraping: HTML/XML to JSON
- Business and Sales
- Finance
- Education
- Healthcare: Diagnose diseases, and generate treatment plans.
- Text Analysis: Sentiment analysis, Entity extraction, and text classification