API Keys & Environment

Relevant source files

The following files were used as context for generating this wiki page:

• config.yaml
• main.py
• LLM.py
• agents.py
• searcher/sementic_search.py
• README.md

The API Keys & Environment module serves as the foundational configuration layer for the CoI-Agent system. It manages the integration of external Large Language Model (LLM) providers and academic search services, ensuring that agents have the necessary credentials and endpoints to perform research tasks, literature searches, and reasoning.

The system supports multiple LLM backends (OpenAI and Azure OpenAI) and utilizes specific environment variables to manage API routing, model selection, and search capabilities through Semantic Scholar.

Sources: config.yaml:1-22, README.md:46-72

Configuration Management

The project uses a config.yaml file to define essential API keys and service endpoints. At runtime, the main.py script parses this file and injects the values into the system's environment variables to be accessed by various modules like LLM.py and agents.py.

Environment Injection Flow

The diagram above shows how configuration data flows from the static YAML file into the runtime environment and subsequent functional modules.
Sources: main.py:10-17, LLM.py:65-103

LLM Provider Integration

CoI-Agent is designed to work with both standard OpenAI and Azure OpenAI services. The choice of provider is determined by the is_azure boolean flag in the configuration.

Supported Providers and Parameters

Parameter	Type	Description	Source
is_azure	Boolean	Toggles between Azure OpenAI and standard OpenAI.	config.yaml:3
MAIN_LLM_MODEL	String	Primary model for complex reasoning (e.g., gpt-4o).	config.yaml:19
CHEAP_LLM_MODEL	String	Secondary model for cost-effective tasks (e.g., gpt-4o-mini).	config.yaml:21
AZURE_OPENAI_ENDPOINT	String	The base URL for the Azure OpenAI resource.	config.yaml:6
OPENAI_BASE_URL	String	Custom base URL for OpenAI (used for proxies).	config.yaml:13

LLM Initialization Logic

The openai_llm class in LLM.py performs a check on environment variables during initialization. If required keys (like AZURE_OPENAI_KEY or OPENAI_API_KEY) are missing or empty strings, the system raises a ValueError.

Sources: LLM.py:66-93, agents.py:16-25

Academic Search Configuration

The system relies on the Semantic Scholar API for literature discovery. This requires a dedicated API key defined in the configuration.

• API Key: SEMENTIC_SEARCH_API_KEY
• Module: SementicSearcher in searcher/sementic_search.py
• Usage: The key is passed in the request headers as x-api-key.

This sequence illustrates the retrieval and application of the search API key during a paper search operation.
Sources: searcher/sementic_search.py:86-90, config.yaml:1

Embedding Services

For semantic ranking and Reranking of papers, the system uses embedding models. These can be configured independently of the main LLM models.

Variable	Usage	Source
EMBEDDING_API_KEY	Credential for the embedding service.	config.yaml:15
EMBEDDING_API_ENDPOINT	Custom endpoint for embedding requests.	config.yaml:16
EMBEDDING_MODEL	The specific model identifier (e.g., text-embedding-3-large).	config.yaml:17

If these are not explicitly set, the system defaults to using the credentials and model settings of the main LLM client.
Sources: LLM.py:126-140, config.yaml:14

Implementation Example: Environment Loading

The following snippet demonstrates how main.py converts YAML configuration into system environment variables:

# main.py:10-17
with open('config.yaml', 'r') as file:
    config = yaml.safe_load(file)
for key, value in config.items():
    if value == "":
        continue
    else:
        os.environ[key] = str(value)

Sources: main.py:10-17

Summary

The API Keys & Environment system provides a flexible, dual-provider architecture. By separating the "Main" and "Cheap" models and centralizing API key management through config.yaml, the project allows for optimized research workflows while maintaining security and ease of setup. Proper configuration of these variables is critical for the ReviewAgent and DeepResearchAgent to function.

Sources: agents.py:15-30, README.md:46-60