valor-lite 0.33.14__py3-none-any.whl → 0.33.16__py3-none-any.whl

valor_lite/text_generation/computation.py

@@ -0,0 +1,609 @@
+import evaluate
+from nltk.tokenize import RegexpTokenizer
+from nltk.translate import bleu_score
+from valor_lite.text_generation.llm.generation import (
+    generate_answer_correctness_verdicts,
+    generate_answer_relevance_verdicts,
+    generate_bias_verdicts,
+    generate_claims,
+    generate_context_precision_verdicts,
+    generate_context_recall_verdicts,
+    generate_context_relevance_verdicts,
+    generate_faithfulness_verdicts,
+    generate_hallucination_verdicts,
+    generate_opinions,
+    generate_statements,
+    generate_summary_coherence,
+    generate_toxicity_verdicts,
+)
+from valor_lite.text_generation.llm.integrations import ClientWrapper
+
+
+def calculate_answer_correctness(
+    client: ClientWrapper,
+    system_prompt: str,
+    query: str,
+    response: str,
+    groundtruths: list[str],
+) -> float:
+    """
+    Compute answer correctness. Answer correctness is computed as an f1 score obtained
+    by comparing prediction statements to ground truth statements.
+
+    If there are multiple ground truths, then the f1 score is computed for each ground
+    truth and the maximum score is returned.
+
+    This metric was adapted from RAGAS. We follow a similar prompting strategy and
+    computation, however we do not do a weighted sum with an answer similarity score
+    using embeddings.
+
+    Parameters
+    ----------
+    client : ClientWrapper
+        The LLM client used to perform evaluation.
+    system_prompt : str
+        A system prompt to pass with the evaluation query.
+    query : str
+        The user query.
+    response : str
+        A generated response.
+    groundtruths : list[str]
+        A list of ground truth references.
+
+    Returns
+    -------
+    float
+        The answer correctness score between 0 and 1. Higher values indicate that the
+        answer is more correct. A score of 1 indicates that all statements in the
+        prediction are supported by the ground truth and all statements in the ground
+        truth are present in the prediction.
+    """
+    prediction_statements = generate_statements(
+        client=client,
+        system_prompt=system_prompt,
+        text=response,
+    )
+    f1_scores = [0.0]
+    for groundtruth in groundtruths:
+        groundtruth_statements = generate_statements(
+            client=client,
+            system_prompt=system_prompt,
+            text=groundtruth,
+        )
+        verdicts = generate_answer_correctness_verdicts(
+            client=client,
+            system_prompt=system_prompt,
+            query=query,
+            groundtruth_statements=groundtruth_statements,
+            prediction_statements=prediction_statements,
+        )
+
+        tp = len(verdicts["TP"])
+        fp = len(verdicts["FP"])
+        fn = len(verdicts["FN"])
+
+        f1_scores.append(tp / (tp + 0.5 * (fp + fn)) if tp > 0 else 0)
+
+    return max(f1_scores)
+
+
+def calculate_answer_relevance(
+    client: ClientWrapper,
+    system_prompt: str,
+    query: str,
+    response: str,
+) -> float:
+    """
+    Compute answer relevance, the proportion of the model response that is
+    relevant to the query, for a single piece of text.
+
+    Parameters
+    ----------
+    client : ClientWrapper
+        The LLM client used to perform evaluation.
+    system_prompt : str
+        A system prompt to pass with the evaluation query.
+    query : str
+        The user query.
+    response : str
+        A generated response.
+
+    Returns
+    -------
+    float
+        The answer relevance score between 0 and 1. A score of 1 indicates that all
+        statements are relevant to the query.
+    """
+    statements = generate_statements(
+        client=client,
+        system_prompt=system_prompt,
+        text=response,
+    )
+    verdicts = generate_answer_relevance_verdicts(
+        client=client,
+        system_prompt=system_prompt,
+        query=query,
+        statements=statements,
+    )
+    if len(verdicts) == 0:
+        return 0.0
+
+    return sum(verdict["verdict"] == "yes" for verdict in verdicts) / len(
+        verdicts
+    )
+
+
+def calculate_bias(
+    client: ClientWrapper,
+    system_prompt: str,
+    response: str,
+) -> float:
+    """
+    Compute bias, the proportion of model opinions that are biased.
+
+    Parameters
+    ----------
+    client : ClientWrapper
+        The LLM client used to perform evaluation.
+    system_prompt : str
+        A system prompt to pass with the evaluation query.
+    response : str
+        A generated response.
+
+    Returns
+    -------
+    float
+        The bias score between 0 and 1. A score of 1 indicates that all opinions in
+        the text are biased.
+    """
+
+    opinions = generate_opinions(
+        client=client,
+        system_prompt=system_prompt,
+        text=response,
+    )
+    if len(opinions) == 0:
+        return 0.0
+
+    verdicts = generate_bias_verdicts(
+        client=client,
+        system_prompt=system_prompt,
+        opinions=opinions,
+    )
+    return sum(verdict["verdict"] == "yes" for verdict in verdicts) / len(
+        verdicts
+    )
+
+
+def calculate_context_precision(
+    client: ClientWrapper,
+    system_prompt: str,
+    query: str,
+    predicted_context: list[str],
+    groundtruth_context: list[str],
+) -> float:
+    """
+    Compute context precision, a score for evaluating the retrieval
+    mechanism of a RAG model.
+
+    First, an LLM is prompted to determine if each context in the context
+    list is useful for producing the ground truth answer to the query.
+
+    If there are multiple ground truths, then the verdict is "yes" for a
+    context if that context is useful for producing any of the ground truth
+    answers, and "no" otherwise.
+
+    Then, using these verdicts, the context precision score is computed as
+    a weighted sum of the precision at k for each k from 1 to the length
+    of the context list.
+
+    Note that the earlier a piece of context appears in the context list,
+    the more important it is in the computation of this score. For example,
+    the first context in the context list will be included in every precision
+    at k computation, so will have a large influence on the final score,
+    whereas the last context will only be used for the last precision at
+    k computation, so will have a small influence on the final score.
+
+    Parameters
+    ----------
+    client : ClientWrapper
+        The LLM client used to perform evaluation.
+    system_prompt : str
+        A system prompt to pass with the evaluation query.
+    query : str
+        The user query.
+    response : str
+        A generated response.
+    predicted_context : list[str]
+        A list of predicted context.
+    groundtruths : list[str]
+        A list of ground truth references.
+
+    Returns
+    -------
+    float
+        The context precision score between 0 and 1. A higher score indicates
+        better context precision.
+    """
+    if len(predicted_context) == 0 and len(groundtruth_context) == 0:
+        return 1.0
+    elif len(predicted_context) == 0 or len(groundtruth_context) == 0:
+        return 0.0
+
+    # Get verdicts for each ground truth, and aggregate by setting the verdict for
+    # a context to "yes" if the verdict is "yes" for any ground truth.
+    aggregate_verdicts = ["no"] * len(predicted_context)
+    for groundtruth in groundtruth_context:
+        verdicts = generate_context_precision_verdicts(
+            client=client,
+            system_prompt=system_prompt,
+            query=query,
+            ordered_context_list=predicted_context,
+            groundtruth=groundtruth,
+        )
+        for i in range(len(verdicts)):
+            if verdicts[i]["verdict"] == "yes":
+                aggregate_verdicts[i] = "yes"
+
+    # Use the aggregate verdicts to compute the precision at k for each k.
+    precision_at_k_list = []
+    for k in range(1, len(predicted_context) + 1):
+        # Only compute the precision at k if the kth context is relevant.
+        if aggregate_verdicts[k - 1] == "yes":
+            precision_at_k = (
+                sum(verdict == "yes" for verdict in aggregate_verdicts[:k]) / k
+            )
+            precision_at_k_list.append(precision_at_k)
+
+    # If none of the context are relevant, then the context precision is 0.
+    if len(precision_at_k_list) == 0:
+        return 0.0
+
+    # Average over all the precision at k for which the kth context is relevant.
+    return sum(precision_at_k_list) / len(precision_at_k_list)
+
+
+def calculate_context_recall(
+    client: ClientWrapper,
+    system_prompt: str,
+    predicted_context: list[str],
+    groundtruth_context: list[str],
+) -> float:
+    """
+    Compute context recall, a score for evaluating the retrieval mechanism of a RAG model.
+
+    The context recall score is the proportion of statements in the ground truth
+    that are attributable to the context list.
+
+    If multiple ground truths are provided, then the context recall score is
+    computed for each ground truth and the maximum score is returned.
+
+    Parameters
+    ----------
+    client : ClientWrapper
+        The LLM client used to perform evaluation.
+    system_prompt : str
+        A system prompt to pass with the evaluation query.
+    predicted_context : list[str]
+        A list of predicted context.
+    groundtruths : list[str]
+        A list of ground truth references.
+
+    Returns
+    -------
+    float
+        The context recall score between 0 and 1. A score of 1 indicates that
+        all ground truth statements are attributable to the contexts in the context list.
+    """
+    if len(predicted_context) == 0 and len(groundtruth_context) == 0:
+        return 1.0
+    elif len(predicted_context) == 0 or len(groundtruth_context) == 0:
+        return 0.0
+
+    scores = []
+    for groundtruth in groundtruth_context:
+        groundtruth_statements = generate_statements(
+            client=client,
+            system_prompt=system_prompt,
+            text=groundtruth,
+        )
+        verdicts = generate_context_recall_verdicts(
+            client=client,
+            system_prompt=system_prompt,
+            context_list=predicted_context,
+            groundtruth_statements=groundtruth_statements,
+        )
+        scores.append(
+            sum(verdict["verdict"] == "yes" for verdict in verdicts)
+            / len(verdicts)
+        )
+
+    return max(scores)
+
+
+def calculate_context_relevance(
+    client: ClientWrapper,
+    system_prompt: str,
+    query: str,
+    context: list[str],
+) -> float:
+    """
+    Compute context relevance, the proportion of contexts in the context list
+    that are relevant to the query.
+
+    Parameters
+    ----------
+    client : ClientWrapper
+        The LLM client used to perform evaluation.
+    system_prompt : str
+        A system prompt to pass with the evaluation query.
+    query : str
+        The user query.
+    context : list[str]
+        A list of predicted context.
+
+    Returns
+    -------
+    float
+        The context relevance score between 0 and 1. A score of 0 indicates
+        that none of the contexts are relevant and a score of 1 indicates
+        that all of the contexts are relevant.
+    """
+    if len(context) == 0:
+        return 0.0
+    verdicts = generate_context_relevance_verdicts(
+        client=client,
+        system_prompt=system_prompt,
+        query=query,
+        context_list=context,
+    )
+    return sum(verdict["verdict"] == "yes" for verdict in verdicts) / len(
+        verdicts
+    )
+
+
+def calculate_faithfulness(
+    client: ClientWrapper,
+    system_prompt: str,
+    response: str,
+    context: list[str],
+) -> float:
+    """
+    Compute the faithfulness score. The faithfulness score is the proportion
+    of claims in the text that are implied by the list of contexts. Claims
+    that contradict the list of contexts and claims that are unrelated to
+    the list of contexts both count against the score.
+
+    Parameters
+    ----------
+    client : ClientWrapper
+        The LLM client used to perform evaluation.
+    system_prompt : str
+        A system prompt to pass with the evaluation query.
+    response : str
+        A generated response.
+    context : list[str]
+        A list of predicted context.
+
+    Returns
+    -------
+    float
+        The faithfulness score between 0 and 1. A score of 1 indicates that
+        all claims in the text are implied by the list of contexts.
+    """
+    if len(context) == 0:
+        return 0.0
+
+    claims = generate_claims(
+        client=client, system_prompt=system_prompt, text=response
+    )
+
+    # If there aren't any claims, then the text is perfectly faithful, as the text does not contain any non-faithful claims.
+    if len(claims) == 0:
+        return 1.0
+
+    faithfulness_verdicts = generate_faithfulness_verdicts(
+        client=client,
+        system_prompt=system_prompt,
+        claims=claims,
+        context_list=context,
+    )
+    return sum(
+        verdict["verdict"] == "yes" for verdict in faithfulness_verdicts
+    ) / len(faithfulness_verdicts)
+
+
+def calculate_hallucination(
+    client: ClientWrapper,
+    system_prompt: str,
+    response: str,
+    context: list[str],
+) -> float:
+    """
+    Compute the hallucination score, the proportion of contexts in the context
+    list that are contradicted by the text.
+
+    Parameters
+    ----------
+    client : ClientWrapper
+        The LLM client used to perform evaluation.
+    system_prompt : str
+        A system prompt to pass with the evaluation query.
+    response : str
+        A generated response.
+    context : list[str]
+        A list of predicted context.
+
+    Returns
+    -------
+    float
+        The hallucination score between 0 and 1. A score of 1 indicates that
+        all contexts are contradicted by the text.
+    """
+    if len(context) == 0:
+        raise ValueError("Hallucination requires context to be calculated.")
+
+    verdicts = generate_hallucination_verdicts(
+        client=client,
+        system_prompt=system_prompt,
+        text=response,
+        context_list=context,
+    )
+    return sum(verdict["verdict"] == "yes" for verdict in verdicts) / len(
+        verdicts
+    )
+
+
+def calculate_summary_coherence(
+    client: ClientWrapper,
+    system_prompt: str,
+    text: str,
+    summary: str,
+) -> int:
+    """
+    Compute summary coherence, the collective quality of a summary.
+
+    Parameters
+    ----------
+    client : ClientWrapper
+        The LLM client used to perform evaluation.
+    system_prompt : str
+        A system prompt to pass with the evaluation query.
+    text : str
+        The original text.
+    summary : str
+        The generated summary.
+
+    Returns
+    -------
+    int
+        The summary coherence score between 1 and 5. A score of 1 indicates
+        the lowest summary coherence and a score of 5 indicates the highest
+        summary coherence.
+    """
+    return generate_summary_coherence(
+        client=client,
+        system_prompt=system_prompt,
+        text=text,
+        summary=summary,
+    )
+
+
+def calculate_toxicity(
+    client: ClientWrapper,
+    system_prompt: str,
+    response: str,
+) -> float:
+    """
+    Compute toxicity, the proportion of opinions that are toxic.
+
+    Parameters
+    ----------
+    client : ClientWrapper
+        The LLM client used to perform evaluation.
+    system_prompt : str
+        A system prompt to pass with the evaluation query.
+    response : str
+        A generated response.
+
+    Returns
+    -------
+    Metric
+        The toxicity score will be evaluated as a float between 0 and 1, with
+        1 indicating that all opinions in the text are toxic.
+    """
+    opinions = generate_opinions(
+        client=client,
+        system_prompt=system_prompt,
+        text=response,
+    )
+    if len(opinions) == 0:
+        return 0.0
+
+    verdicts = generate_toxicity_verdicts(
+        client=client,
+        system_prompt=system_prompt,
+        opinions=opinions,
+    )
+    return sum(verdict["verdict"] == "yes" for verdict in verdicts) / len(
+        verdicts
+    )
+
+
+def calculate_rouge_scores(
+    prediction: str,
+    references: str | list[str],
+    rouge_types: list[str],
+    use_stemmer: bool = False,
+) -> dict[str, float]:
+    """
+    Calculate ROUGE scores for a prediction given some set of references.
+
+    Parameters
+    ----------
+    prediction : str
+        A generated response to score. Each prediction should be a string with tokens separated by spaces.
+    references : str | list[str]
+        A list of references for a given response. Each reference should be a string with tokens separated by spaces.
+    rouge_types : list[str]
+        A list of rouge types to calculate.
+    use_stemmer: bool, default=False
+        If True, uses Porter stemmer to strip word suffixes. Defaults to False.
+    """
+    rouge = evaluate.load("rouge")
+
+    metrics = rouge.compute(
+        predictions=[prediction],
+        references=[references],
+        rouge_types=rouge_types,
+        use_stemmer=use_stemmer,
+        use_aggregator=False,  # aggregation gives us an average across all predictions, which isn't what we want
+    )
+
+    # find the max value for each prediction
+    results = dict()
+    if metrics is not None:
+        for type_ in rouge_types:
+            results[type_] = max(metrics[type_][0], 0.0)
+    return results
+
+
+def calculate_sentence_bleu(
+    prediction: str,
+    references: list[str],
+    weights: tuple[float, ...] | list[float],
+) -> float:
+    """
+    Calculate sentence BLEU scores for a of prediction - ground truth pair.
+
+    Parameters
+    ----------
+    prediction : str
+        A generated response to score. Each prediction should be a string with tokens separated by spaces.
+    references : list[str]
+        A list of references for a given response. Each reference should be a string with tokens separated by spaces.
+    weights : tuple[float]
+        The default BLEU calculates a score for up to 4-grams using uniform
+        weights (this is called BLEU-4). To evaluate your translations with
+        higher/lower order ngrams, use customized weights. Example: when accounting
+        for up to 5-grams with uniform weights (this is called BLEU-5) use [1/5]*5
+    """
+    if len(weights) == 0:
+        raise ValueError("At least one weight should be defined.")
+
+    tokenizer = RegexpTokenizer(
+        r"\w+|\$[\d]+|[^\s\.]+"
+    )  # regex tokenizer that ignores periods
+
+    tokenized_prediction = tokenizer.tokenize(prediction)
+    tokenized_references = [tokenizer.tokenize(ref) for ref in references]
+
+    # find the max value for each prediction
+    result = float(
+        bleu_score.sentence_bleu(
+            references=tokenized_references,
+            hypothesis=tokenized_prediction,
+            weights=weights,
+        ),  # type: ignore
+    )
+    return result if result >= 1e-9 else 0.0

valor_lite/text_generation/llm/exceptions.py

@@ -0,0 +1,14 @@
+class InvalidLLMResponseError(Exception):
+    """
+    Raised when the response from the LLM is invalid for a given metric computation.
+    """
+
+    pass
+
+
+class MismatchingTextGenerationDatumError(Exception):
+    """
+    Raised when datums with the same uid but different text are added to the ValorTextGenerationStreamingManager.
+    """
+
+    pass