R&D Agent: Focusing on automating the most core and valuable part of the industrial R&D process.
Core method: Evolving;
# Scenarios
## Data-driven R&D
TODO: importance justification
### 🎯 Goal
In this project, we are aiming to build a Data-Centric R\&D Agent that can
+ Read real-world material (reports, papers, etc.) and extract key formulas, descriptions of interested features, factors and models.
+ Implement the extracted formulas, features, factors and models in runnable codes.
+ Due the limited ability for LLM in implementing in once, evolving the agent to be able to extend abilities by learn from feedback and knowledge and improve the agent's ability to implement more complex models.
+ Further propose new ideas based on current knowledge and observations.
In this section, we will briefly introduce the roadmap/technical type of this project.
1. Backbone LLM: We use GPT series as main backbone of the agent. `.env` file uis used to config settings (such as APIkey, APIEndpoint and etc) in the environment variables way. Check this [Readme](src/scripts/benchmark/README.md) for environment set up.
2. KnowledgeGraph based evolving: We do not do any further pertain or fine-tune on the LLM model. Instead, we modify prompts like RAG, but use knowledge graph query information to evolve the agent's ability to implement more complex models.
+ Typically, we build a knowledge consisted with `Error`, `Component`(you can think of it as a numeric operation or function), `Trail` and etc. We add nodes of these types to the knowledge graph with relationship while the agent tries to implement a model. For each attempts, the agent will query the knowledge graph to get the information of current status as prompt input. The agent will also update the knowledge graph with the new information after the attempt.
## Code Refinement
Example: code standard, design. Lint
# 🔧 Development
- Set up the development environment.
```bash
make dev
```
- Run linting and formatting.
```bash
make lint
```
### 📚 Code Structure
1. Backbone/APIBackend of LLm are encapsulated in [src/finco/llm.py](src/finco/llm.py). All chat completion request are managed by this file.
2. All frequently modified codes under tense development are included in the [src/scripts](src/scripts) folder.
+ The most important task is to improve the agent's performance in the benchmark of factor implementation.
3. Currently, factor implementation is the main task. We define basic class of factor implementation in [src/scripts/factor_implementation/share_modules] and implementation strategies in [src/scripts/factor_implementation/baselines].
### 🔮 Future Code Structure
Currently, the code structure is unstable and will frequently change for quick updates. The code will be refactored before a standard release. Please try to align with the following principles when developing to minimize the effort required for future refactoring.
| 📁 core | The core framework of the system. All classes should be abstract and usually can't be used directly. |
| 📁 component X | Useful components that can be used by others(e.g. scenario). Many subclasses of core classes are located here. |
| 📁 app | Applications for specific scenarios (usually built based on components). Removing any of them does not affect the system's completeness or other scenarios. |
| 📁 scripts | Quick and dirty things. These are candidates for core, components, and apps. |
This project welcomes contributions and suggestions.
You can find issues in the issues list or simply running `grep -r "TODO:"`.
Making contributions is not a hard thing. Solving an issue(maybe just answering a question raised in issues list ), fixing/issuing a bug, improving the documents and even fixing a typo are important contributions to RDAgent.
