haystack-experimental 0.2.0

Experimental components and features for the Haystack LLM framework.

Haystack experimental package

The haystack-experimental package provides Haystack users with access to experimental features without immediately committing to their official release. The main goal is to gather user feedback and iterate on new features quickly.

Installation

For simplicity, every release of haystack-experimental will ship all the available experiments at that time. To install the latest experimental features, run:

$ pip install -U haystack-experimental

[!IMPORTANT] The latest version of the experimental package is only tested against the latest version of Haystack. Compatibility with older versions of Haystack is not guaranteed.

Experiments lifecycle

Each experimental feature has a default lifespan of 3 months starting from the date of the first non-pre-release build that includes it. Once it reaches the end of its lifespan, the experiment will be either:

• Merged into Haystack core and published in the next minor release, or
• Released as a Core Integration, or
• Dropped.

Experiments catalog

The latest version of the package contains the following experiments:

Name	Type	Expected End Date	Dependencies	Cookbook	Discussion
EvaluationHarness	Evaluation orchestrator	October 2024	None		Discuss
OpenAIFunctionCaller	Function Calling Component	October 2024	None	🔜
OpenAPITool	OpenAPITool component	October 2024	jsonref		Discuss
Support for Tools: refactored ChatMessage dataclass, Tool dataclass, refactored OpenAIChatGenerator, ToolInvoker component	Tool Calling support	November 2024	jsonschema		Discuss
ChatMessageWriter	Memory Component	December 2024	None		Discuss
ChatMessageRetriever	Memory Component	December 2024	None		Discuss
InMemoryChatMessageStore	Memory Store	December 2024	None		Discuss
Auto-Merging Retriever & HierarchicalDocumentSplitter	Document Splitting & Retrieval Technique	December 2024	None		Discuss

Usage

Experimental new features can be imported like any other Haystack integration package:

from haystack.dataclasses import ChatMessage
from haystack_experimental.components.generators import FoobarGenerator

c = FoobarGenerator()
c.run([ChatMessage.from_user("What's an experiment? Be brief.")])

Experiments can also override existing Haystack features. For example, users can opt into an experimental type of Pipeline by just changing the usual import:

# from haystack import Pipeline
from haystack_experimental import Pipeline

pipe = Pipeline()
# ...
pipe.run(...)

Some experimental features come with example notebooks and resources that can be found in the examples folder.

Documentation

Documentation for haystack-experimental can be found here.

Implementation

Experiments should replicate the namespace of the core package. For example, a new generator:

# in haystack_experimental/components/generators/foobar.py

from haystack import component


@component
class FoobarGenerator:
    ...

When the experiment overrides an existing feature, the new symbol should be created at the same path in the experimental package. This new symbol will override the original in haystack-ai: for classes, with a subclass and for bare functions, with a wrapper. For example:

# in haystack_experiment/src/haystack_experiment/core/pipeline/pipeline.py

from haystack.core.pipeline import Pipeline as HaystackPipeline


class Pipeline(HaystackPipeline):
    # Any new experimental method that doesn't exist in the original class
    def run_async(self, inputs) -> Dict[str, Dict[str, Any]]:
        ...

    # Existing methods with breaking changes to their signature, like adding a new mandatory param
    def to_dict(new_param: str) -> Dict[str, Any]:
        # do something with the new parameter
        print(new_param)
        # call the original method
        return super().to_dict()

Contributing

Direct contributions to haystack-experimental are not expected, but Haystack maintainers might ask contributors to move pull requests that target the core repository to this repository.

Telemetry

As with the Haystack core package, we rely on anonymous usage statistics to determine the impact and usefulness of the experimental features. For more information on what we collect and how we use the data, as well as instructions to opt-out, please refer to our documentation.