trulens-feedback 1.1.0a2

A TruLens extension package implementing feedback functions for LLM App evaluation.

trulens-feedback

Feedback Functions

The Feedback class contains the starting point for feedback function specification and evaluation. A typical use-case looks like this:

from trulens.core import Feedback, Select, Feedback

hugs = feedback.Huggingface()

f_lang_match = Feedback(hugs.language_match)
    .on_input_output()

The components of this specifications are:

• Provider classes -- feedback.OpenAI contains feedback function implementations like context_relevance. Other classes subtyping feedback.Provider include Huggingface and Cohere.

• Feedback implementations -- provider.context_relevance is a feedback function implementation. Feedback implementations are simple callables that can be run on any arguments matching their signatures. In the example, the implementation has the following signature:

  def language_match(self, text1: str, text2: str) -> float:
  

  That is, language_match is a plain python method that accepts two pieces of text, both strings, and produces a float (assumed to be between 0.0 and 1.0).

• Feedback constructor -- The line Feedback(provider.language_match) constructs a Feedback object with a feedback implementation.

• Argument specification -- The next line, on_input_output, specifies how the language_match arguments are to be determined from an app record or app definition. The general form of this specification is done using on but several shorthands are provided. on_input_output states that the first two argument to language_match (text1 and text2) are to be the main app input and the main output, respectively.

  Several utility methods starting with .on provide shorthands:

  • on_input(arg) == on_prompt(arg: Optional[str]) -- both specify that the next unspecified argument or arg should be the main app input.

  • on_output(arg) == on_response(arg: Optional[str]) -- specify that the next argument or arg should be the main app output.

  • on_input_output() == on_input().on_output() -- specifies that the first two arguments of implementation should be the main app input and main app output, respectively.

  • on_default() -- depending on signature of implementation uses either on_output() if it has a single argument, or on_input_output if it has two arguments.

  Some wrappers include additional shorthands:

llama_index-specific selectors

• TruLlama.select_source_nodes() -- outputs the selector for the source documents part of the engine output.
• TruLlama.select_context() -- outputs the selector for the text of the source documents part of the engine output.

langchain-specific selectors

• Langchain.select_context() -- outputs the selector for retrieved context from the app's internal get_relevant_documents method.

NeMo-specific selectors

• NeMo.select_context() -- outputs the selector for the retrieved context from the app's internal search_relevant_chunks method.

Fine-grained Selection and Aggregation

For more advanced control on the feedback function operation, we allow data selection and aggregation. Consider this feedback example:

f_context_relevance = Feedback(openai.context_relevance)
    .on_input()
    .on(Select.Record.app.combine_docs_chain._call.args.inputs.input_documents[:].page_content)
    .aggregate(numpy.min)

# Implementation signature:
# def context_relevance(self, question: str, statement: str) -> float:

• Argument Selection specification -- Where we previously set, on_input_output , the on(Select...) line enables specification of where the statement argument to the implementation comes from. The form of the specification will be discussed in further details in the Specifying Arguments section.

• Aggregation specification -- The last line aggregate(numpy.min) specifies how feedback outputs are to be aggregated. This only applies to cases where the argument specification names more than one value for an input. The second specification, for statement was of this type. The input to aggregate must be a method which can be imported globally. This requirement is further elaborated in the next section. This function is called on the float results of feedback function evaluations to produce a single float. The default is numpy.mean.

The result of these lines is that f_context_relevance can be now be run on app/records and will automatically select the specified components of those apps/records:

record: Record = ...
app: App = ...

feedback_result: FeedbackResult = f_context_relevance.run(app=app, record=record)

The object can also be provided to an app wrapper for automatic evaluation:

app: App = TruChain(...., feedbacks=[f_context_relevance])

Specifying Implementation Function and Aggregate

The function or method provided to the Feedback constructor is the implementation of the feedback function which does the actual work of producing a float indicating some quantity of interest.

Note regarding FeedbackMode.DEFERRED -- Any function or method (not static or class methods presently supported) can be provided here but there are additional requirements if your app uses the "deferred" feedback evaluation mode (when feedback_mode=FeedbackMode.DEFERRED are specified to app constructor). In those cases the callables must be functions or methods that are importable (see the next section for details). The function/method performing the aggregation has the same requirements.

Import requirement (DEFERRED feedback mode only)

If using deferred evaluation, the feedback function implementations and aggregation implementations must be functions or methods from a Provider subclass that is importable. That is, the callables must be accessible were you to evaluate this code:

from somepackage.[...] import someproviderclass
from somepackage.[...] import somefunction

# [...] means optionally further package specifications

provider = someproviderclass(...) # constructor arguments can be included
feedback_implementation1 = provider.somemethod
feedback_implementation2 = somefunction

For provided feedback functions, somepackage is trulens.feedback and someproviderclass is OpenAI or one of the other Provider subclasses. Custom feedback functions likewise need to be importable functions or methods of a provider subclass that can be imported. Critically, functions or classes defined locally in a notebook will not be importable this way.

Specifying Arguments

The mapping between app/records to feedback implementation arguments is specified by the on... methods of the Feedback objects. The general form is:

feedback: Feedback = feedback.on(argname1=selector1, argname2=selector2, ...)

That is, Feedback.on(...) returns a new Feedback object with additional argument mappings, the source of argname1 is selector1 and so on for further argument names. The types of selector1 is JSONPath which we elaborate on in the "Selector Details".

If argument names are omitted, they are taken from the feedback function implementation signature in order. That is,

Feedback(...).on(argname1=selector1, argname2=selector2)

and

Feedback(...).on(selector1, selector2)

are equivalent assuming the feedback implementation has two arguments, argname1 and argname2, in that order.

Running Feedback

Feedback implementations are simple callables that can be run on any arguments matching their signatures. However, once wrapped with Feedback, they are meant to be run on outputs of app evaluation (the "Records"). Specifically, Feedback.run has this definition:

def run(self,
    app: Union[AppDefinition, JSON],
    record: Record
) -> FeedbackResult:

That is, the context of a Feedback evaluation is an app (either as AppDefinition or a JSON-like object) and a Record of the execution of the aforementioned app. Both objects are indexable using "Selectors". By indexable here we mean that their internal components can be specified by a Selector and subsequently that internal component can be extracted using that selector. Selectors for Feedback start by specifying whether they are indexing into an App or a Record via the __app__ and __record__ special attributes (see Selectors section below).

Selector Details

Selectors are of type JSONPath defined in util.py but are also aliased in schema.py as Select.Query. Objects of this type specify paths into JSON-like structures (enumerating Record or App contents).

By JSON-like structures we mean python objects that can be converted into JSON or are base types. This includes:

• base types: strings, integers, dates, etc.

• sequences

• dictionaries with string keys

Additionally, JSONPath also index into general python objects like AppDefinition or Record though each of these can be converted to JSON-like.

When used to index json-like objects, JSONPath are used as generators: the path can be used to iterate over items from within the object:

class JSONPath...
    ...
    def __call__(self, obj: Any) -> Iterable[Any]:
    ...

In most cases, the generator produces only a single item but paths can also address multiple items (as opposed to a single item containing multiple).

The syntax of this specification mirrors the syntax one would use with instantiations of JSON-like objects. For every obj generated by query: JSONPath:

• query[somekey] generates the somekey element of obj assuming it is a dictionary with key somekey.

• query[someindex] generates the index someindex of obj assuming it is a sequence.

• query[slice] generates the multiple elements of obj assuming it is a sequence. Slices include : or in general startindex:endindex:step.

• query[somekey1, somekey2, ...] generates multiple elements of obj assuming obj is a dictionary and somekey1... are its keys.

• query[someindex1, someindex2, ...] generates multiple elements indexed by someindex1... from a sequence obj.

• query.someattr depends on type of obj. If obj is a dictionary, then query.someattr is an alias for query[someattr]. Otherwise if someattr is an attribute of a python object obj, then query.someattr generates the named attribute.

For feedback argument specification, the selectors should start with either __record__ or __app__ indicating which of the two JSON-like structures to select from (Records or Apps). Select.Record and Select.App are defined as Query().__record__ and Query().__app__ and thus can stand in for the start of a selector specification that wishes to select from a Record or App, respectively. The full set of Query aliases are as follows:

• Record = Query().__record__ -- points to the Record.

• App = Query().app -- points to the App.

• RecordInput = Record.main_input -- points to the main input part of a Record. This is the first argument to the root method of an app (for langchain Chains this is the __call__ method).

• RecordOutput = Record.main_output -- points to the main output part of a Record. This is the output of the root method of an app (i.e. __call__ for langchain Chains).

• RecordCalls = Record.app -- points to the root of the app-structured mirror of calls in a record. See App-organized Calls Section above.

Multiple Inputs Per Argument

As in the f_context_relevance example, a selector for a single argument may point to more than one aspect of a record/app. These are specified using the slice or lists in key/index positions. In that case, the feedback function is evaluated multiple times, its outputs collected, and finally aggregated into a main feedback result.

The collection of values for each argument of feedback implementation is collected and every combination of argument-to-value mapping is evaluated with a feedback definition. This may produce a large number of evaluations if more than one argument names multiple values. In the dashboard, all individual invocations of a feedback implementation are shown alongside the final aggregate result.

App/Record Organization (What can be selected)

Apps are serialized into JSON-like structures which are indexed via selectors. The exact makeup of this structure is app-dependent though always start with app, that is, the trulens wrappers (subtypes of App) contain the wrapped app in the attribute app:

# app.py:
class App(AppDefinition, SerialModel):
    ...
    # The wrapped app.
    app: Any = Field(exclude=True)
    ...

For your app, you can inspect the JSON-like structure by using the dict method:

app = ... # your app, extending App
print(app.dict())

The other non-excluded fields accessible outside of the wrapped app are listed in the AppDefinition class in schema.py:

class AppDefinition(WithClassInfo, SerialModel, ABC):
    ...

    app_id: AppID

    feedback_definitions: Sequence[FeedbackDefinition] = []

    feedback_mode: FeedbackMode = FeedbackMode.WITH_APP_THREAD

    root_class: Class

    root_callable: ClassVar[FunctionOrMethod]

    app: JSON

Note that app is in both classes. This distinction between App and AppDefinition here is that one corresponds to potentially non-serializable python objects (App) and their serializable versions (AppDefinition). Feedbacks should expect to be run with AppDefinition. Fields of App that are not part of AppDefinition may not be available.

You can inspect the data available for feedback definitions in the dashboard by clicking on the "See full app json" button on the bottom of the page after selecting a record from a table.

The other piece of context to Feedback evaluation are records. These contain the inputs/outputs and other information collected during the execution of an app:

class Record(SerialModel):
    record_id: RecordID
    app_id: AppID

    cost: Optional[Cost] = None
    perf: Optional[Perf] = None

    ts: datetime = pydantic.Field(default_factory=lambda: datetime.now())

    tags: str = ""

    main_input: Optional[JSON] = None
    main_output: Optional[JSON] = None  # if no error
    main_error: Optional[JSON] = None  # if error

    # The collection of calls recorded. Note that these can be converted into a
    # json structure with the same paths as the app that generated this record
    # via `layout_calls_as_app`.
    calls: Sequence[RecordAppCall] = []

A listing of a record can be seen in the dashboard by clicking the "see full record json" button on the bottom of the page after selecting a record from the table.

Calls made by App Components

When evaluating a feedback function, Records are augmented with app/component calls in app layout in the attribute app. By this we mean that in addition to the fields listed in the class definition above, the app field will contain the same information as calls but organized in a manner mirroring the organization of the app structure. For example, if the instrumented app contains a component combine_docs_chain then app.combine_docs_chain will contain calls to methods of this component. In the example at the top of this docstring, _call was an example of such a method. Thus app.combine_docs_chain._call further contains a RecordAppCall (see schema.py) structure with information about the inputs/outputs/metadata regarding the _call call to that component. Selecting this information is the reason behind the Select.RecordCalls alias (see next section).

You can inspect the components making up your app via the App method print_instrumented.