Creating Abstractions

OpenBench uses abstract base classes for extensibility.

DataSource

Extract data from any source:

from openbench.core import DataSource, RawData

class MyDataSource(DataSource):
    def __init__(self, config: str):
        self.config = config

    @property
    def source_type(self) -> str:
        return "my-source"

    @property
    def source_id(self) -> str:
        return f"my-source:{self.config}"

    def get_metadata(self) -> dict:
        return {"type": self.source_type, "config": self.config}

    def extract(self) -> RawData:
        content = self._fetch_content()
        return RawData(
            content=content,
            content_type="text",  # or "structured", "binary"
            metadata=self.get_metadata(),
            source=self
        )

    def validate(self) -> bool:
        return True

Agent

Execute AI tasks:

from openbench.core import Agent, ExecutionContext, ExecutionResult

class MyAgent(Agent):
    def __init__(self, goal: str, model: str = "gemini-2.5-flash"):
        self.goal = goal
        self.model = model

    @property
    def agent_type(self) -> str:
        return "custom"

    def execute(self, context: ExecutionContext) -> ExecutionResult:
        output = self._process(context.input_data)
        return ExecutionResult(
            output=output,
            status="completed",
            metadata={"model": self.model},
            cost=0.05,
            tokens_used=500
        )

    def estimate_cost(self, context: ExecutionContext) -> float:
        return 0.05

OutputGenerator

Generate output artifacts:

from openbench.core import OutputGenerator, GeneratedOutput

class MyOutputGenerator(OutputGenerator):
    def __init__(self, template: str = "default"):
        self.template = template

    @property
    def output_format(self) -> str:
        return "custom"

    def generate(self, content, template=None, **options) -> GeneratedOutput:
        file_path = self._create_file(content, template or self.template)
        return GeneratedOutput(
            file_path=file_path,
            format=self.output_format,
            size_bytes=self._get_file_size(file_path),
            metadata={"template": template, **options}
        )

    def validate(self, content) -> bool:
        return content is not None

DataStore

Store and retrieve data with vector search:

from openbench.core import DataStore
from openbench.data.stores.base import EmbeddingMixin

class MyVectorStore(DataStore, EmbeddingMixin):
    def __init__(self, index_name: str, embedding_model: str = "text-embedding-3-small"):
        self.index_name = index_name
        self._embedding_model = embedding_model

    @property
    def store_type(self) -> str:
        return "my-vector-store"

    def index(self, data, **options) -> str:
        vector = self._embed(data.content)
        # Store vector and return ID
        return item_id

    def search(self, query: str, top_k: int = 5):
        query_vector = self._embed(query)
        # Search and return results
        return results

    def get(self, item_id: str):
        # Retrieve by ID
        pass

    def delete(self, item_id: str) -> bool:
        # Delete by ID
        return True

FrameworkAdapter

Wrap external AI frameworks:

from openbench.core import FrameworkAdapter

class MyFrameworkAdapter(FrameworkAdapter):
    def __init__(self, agent):
        self.agent = agent

    @property
    def framework_name(self) -> str:
        return "my-framework"

    def invoke(self, input, config=None):
        return self.agent.run(input)

    async def ainvoke(self, input, config=None):
        return await self.agent.arun(input)

Registry Pattern

Register and create implementations:

from openbench.core import DataSourceRegistry, AgentRegistry

# Register
DataSourceRegistry.register('custom', 'my-impl', MyDataSource)
AgentRegistry.register('custom', 'my-impl', MyAgent)

# Create (swappable!)
source = DataSourceRegistry.create('custom', 'my-impl', config="value")
agent = AgentRegistry.create('custom', 'my-impl', goal="analyze")

Anti-Patterns

DO NOT:

• Skip abstract properties (source_type, agent_type, output_format, framework_name) - they're required
• Return raw strings from extract() - always return RawData dataclass
• Return raw dicts from execute() - always return ExecutionResult dataclass
• Return raw paths from generate() - always return GeneratedOutput dataclass
• Forget validate() - it's called before processing in the pipeline
• Import external SDKs at module level - use lazy imports with helpful error messages
• Guess method signatures - read src/openbench/core/abstractions.py first

Cross-References

• Composing Workflows: All abstractions are Chainable → see composing-workflows skill
• Data Layer: DataSource and DataStore implementations → see data-layer skill
• Intelligence Layer: Agent and LLMProvider implementations → see intelligence-layer skill
• Output Layer: OutputGenerator implementations → see output-layer skill
• Adapters: FrameworkAdapter implementations → see adapters skill
• Testing: Test patterns for abstractions → see testing-openbench skill

Requirements

1. Always implement required abstract properties
2. Use type hints
3. Handle errors gracefully
4. Write tests for every implementation
5. Document with docstrings

For examples, see src/openbench/data/sources/ and src/openbench/intelligence/agents.py