presentation-search / src /eval /eval_mlflow.py
Ilia Tambovtsev
feat: add presentationidx metric
e887b28
import asyncio
import logging
import os
import time
from abc import ABC, abstractmethod
from datetime import datetime
from json import load
from pathlib import Path
from tempfile import NamedTemporaryFile
from textwrap import dedent
from typing import Any, Dict, List, Optional, Protocol, Union
import mlflow
import mlflow.config
import numpy as np
import pandas as pd
from dotenv import load_dotenv
from langchain_core.output_parsers import JsonOutputParser, StrOutputParser
from langchain_core.prompts import PromptTemplate
from langchain_openai import ChatOpenAI
from pydantic import BaseModel, ConfigDict, Field
from tqdm import tqdm
from tqdm.asyncio import tqdm_asyncio
from src.config import Config, load_spreadsheet
from src.config.logging import setup_logging
from src.config.spreadsheets import (
GoogleSpreadsheetManager,
GoogleSpreadsheetManagerMLFlow,
)
from src.rag import (
ChromaSlideStore,
HyperbolicScorer,
MinScorer,
PresentationRetriever,
ScorerTypes,
)
from src.rag.storage import LLMPresentationRetriever, RetrieverTypes
logger = logging.getLogger(__name__)
class MetricResult(BaseModel):
"""Container for metric calculation results"""
name: str
score: float
explanation: Optional[str] = None
model_config = ConfigDict(arbitrary_types_allowed=True)
class BaseMetric(BaseModel):
"""Base class for evaluation metrics"""
model_config = ConfigDict(arbitrary_types_allowed=True)
@property
def name(self) -> str:
"""Get metric name"""
return self.__class__.__name__.lower()
@abstractmethod
async def acalculate(self, run_output: Dict, ground_truth: Dict) -> MetricResult:
"""Calculate metric value asynchronously"""
pass
def calculate(self, run_output: Dict, ground_truth: Dict) -> MetricResult:
"""Synchronous wrapper for calculate"""
return asyncio.run(self.acalculate(run_output, ground_truth))
class PresentationMatch(BaseMetric):
"""Check if top-1 retrieved presentation matches ground truth"""
async def acalculate(self, run_output: Dict, ground_truth: Dict) -> MetricResult:
best_pres_info = run_output["contexts"][0]
best_pres_name = best_pres_info["pres_name"]
score = float(best_pres_name == ground_truth["pres_name"])
return MetricResult(
name=self.name,
score=score,
explanation=f"Retrieved: {best_pres_name}, Expected: {ground_truth['pres_name']}",
)
class PresentationFound(BaseMetric):
"""Check if ground truth presentation is in top-k"""
async def acalculate(self, run_output: Dict, ground_truth: Dict) -> MetricResult:
found_pres_names = [c["pres_name"] for c in run_output["contexts"]]
score = float(ground_truth["pres_name"] in found_pres_names)
return MetricResult(
name=self.name,
score=score,
explanation=f"Found in positions: {[i for i, p in enumerate(found_pres_names) if p == ground_truth['pres_name']]}",
)
class PresentationIdx(BaseMetric):
async def acalculate(self, run_output: Dict, ground_truth: Dict) -> MetricResult:
found_pres_names = [c["pres_name"] for c in run_output["contexts"]]
score = float("nan")
for i, pres in enumerate(found_pres_names):
if pres == ground_truth["pres_name"]:
score = float(i + 1)
return MetricResult(
name=self.name,
score=score,
explanation=(
f"Presentation was found at position {score}"
if score != float("nan")
else "Presentation was not found"
),
)
class PageMatch(BaseMetric):
"""Check if best page matches ground truth"""
async def acalculate(self, run_output: Dict, ground_truth: Dict) -> MetricResult:
score = 0.0
explanation = ""
for pres_info in run_output["contexts"]:
best_page_found = pres_info["pages"][0]
if pres_info["pres_name"] == ground_truth["pres_name"]:
reference_pages = ground_truth["pages"]
if not reference_pages:
score = 1.0
explanation = "No specific page required"
elif best_page_found in reference_pages:
score = 1.0
explanation = f"Found correct page {best_page_found}"
else:
explanation = f"Page mismatch: found {best_page_found}, expected {reference_pages}"
return MetricResult(name=self.name, score=score, explanation=explanation)
class PageFound(BaseMetric):
"""Check if any of ground truth pages are found in retrieved results
The page is considered found if it appears in ANY position in the correct presentation.
This is less strict than PageMatch which checks best matching page.
"""
async def acalculate(self, run_output: Dict, ground_truth: Dict) -> MetricResult:
"""Calculate metric value"""
score = 0.0
explanation = ""
# Get all pages from each presentation's results
for pres_info in run_output["contexts"]:
# Only check pages from the correct presentation
if pres_info["pres_name"] == ground_truth["pres_name"]:
found_pages = pres_info["pages"]
reference_pages = ground_truth["pages"]
# Handle case when no specific page required
if not reference_pages:
score = 1.0
explanation = "No specific page required"
break
# Check if any reference page is found
matching_pages = set(found_pages) & set(reference_pages)
if matching_pages:
score = 1.0
explanation = f"Found pages {matching_pages} in positions {[found_pages.index(p)+1 for p in matching_pages]}"
break
else:
explanation = f"No matching pages found. Retrieved: {found_pages}, Expected: {reference_pages}"
return MetricResult(name=self.name, score=score, explanation=explanation)
class PresentationCount(BaseMetric):
"""Count number of retrieved presentations"""
async def acalculate(self, run_output: Dict, ground_truth: Dict) -> MetricResult:
"""Count presentations in retrieved results"""
n_pres = len(run_output["contexts"])
return MetricResult(
name=self.name,
score=float(n_pres),
explanation=f"Retrieved {n_pres} presentations",
)
class BestChunkMatch(BaseMetric):
async def acalculate(self, run_output: Dict, ground_truth: Dict) -> MetricResult:
"""Count presentations in retrieved results"""
best_pres = run_output["contexts"][0]
best_chunk = best_pres["best_chunk"]
true_content_type = ground_truth["content_type"]
found_content_type = best_chunk["chunk_type"]
score = 0
if true_content_type in found_content_type: # text_content and visual_content
score = 1
if true_content_type == "general" and found_content_type in [
"general_description",
"conclusions_and_insights",
"layout_and_composition",
]:
score = 1
return MetricResult(
name=self.name,
score=float(score),
explanation=f"Found content type '{found_content_type}' matches ground truth '{true_content_type}'",
)
class LLMRelevance(BaseMetric):
"""LLM-based relevance scoring"""
class RelevanceOutput(BaseModel):
explanation: str = Field(
description="Detailed explanation of why the content is/isn't relevant and how it relates to the query"
)
relevance_score: int = Field(description="Relevance score from 0-10")
model_config = ConfigDict(arbitrary_types_allowed=True)
llm: ChatOpenAI = Field(description="LLM for relevance scoring")
n_contexts: int = Field(default=-1, description="Number of contexts to evaluate")
rate_limit_timeout: float = Field(
default=-1.0, description="Rate limit timeout in seconds"
)
def model_post_init(self, __context: Any):
# fmt: off
prompt_template = PromptTemplate.from_template(dedent(
"""\
You are an expert relevance assessor for a presentation retrieval system. Your task is to evaluate whether the retrieved slide descriptions contain relevant information that answers the user's query. Analyze all provided slide descriptions as a collective unit and provide a detailed explanation along with a relevance score.
Each slide description contains these equally weighted sections:
- Text Content: The actual text present on the slide
- Visual Content: Description of images, charts, or other visual elements
- Topic Overview: Main themes and subjects covered
- Insights and Conclusions: Key takeaways and conclusions
- Layout and Composition: Structural organization of the slide
Scoring Guidelines:
- 9-10: Perfect match - Content directly and comprehensively answers the query (e.g., query asks about sales trends, and slides show exact sales data and analysis)
- 7-8: Strong relevance - Content clearly relates to the query but may miss minor details (e.g., query asks about complete workflow, slides show most but not all steps)
- 5-6: Moderate relevance - Content addresses the query partially or indirectly (e.g., query asks about specific feature, slides discuss it briefly among other topics)
- 3-4: Weak relevance - Content touches the topic but doesn't provide substantial answer (e.g., query asks about implementation details, slides only mention the concept)
- 1-2: Minimal relevance - Only slight connection to the query (e.g., query asks about specific metric, slides only mention related general category)
- 0: No relevance - Content has no connection to the query
Evaluation Rules:
1. Award points if ANY section (text, visual, etc.) contains relevant information
2. In your explanation, cite specific sections and content that justify your score
3. Treat all sections equally - a match in visual content is as valuable as a match in text content
4. Consider all slides collectively - relevant information might be spread across multiple slides
5. Partial matches are valuable if they provide any useful information related to the query
# Slide Descriptions
{context_str}
--- END OF SLIDE DESCRIPTIONS ---
Question: {query_str}
Output formatting:
{format_instructions}
"""))
# fmt: on
self._parser = JsonOutputParser(pydantic_object=self.RelevanceOutput)
self._chain = prompt_template | self.llm.with_structured_output(
self.RelevanceOutput
)
async def acalculate(self, run_output: Dict, ground_truth: Dict) -> MetricResult:
"""Evaluate relevance of retrieved content"""
if self.rate_limit_timeout > 0:
time.sleep(1.05) # Rate limiting
question = ground_truth["question"]
pres = run_output["contexts"][0]
contexts_used = (
pres["contexts"]
if self.n_contexts <= 0
else pres["contexts"][: self.n_contexts]
)
pres_context = "\n\n---\n\n".join(contexts_used)
llm_out = await self._chain.ainvoke(
dict(
query_str=question,
context_str=pres_context,
format_instructions=self._parser.get_format_instructions(),
)
)
llm_out_dict = llm_out.model_dump()
return MetricResult(
name=self.name,
score=float(llm_out_dict["relevance_score"]),
explanation=llm_out_dict["explanation"],
)
class MetricsRegistry:
"""Factory for creating metric instances"""
_metrics = {
"presentationmatch": PresentationMatch,
"presentationfound": PresentationFound,
"presentationidx": PresentationIdx,
"pagematch": PageMatch,
"pagefound": PageFound,
"presentationcount": PresentationCount,
"bestchunkmatch": BestChunkMatch,
"llmrelevance": LLMRelevance,
}
@classmethod
def create(cls, metric_name: str, **kwargs) -> BaseMetric:
"""Create metric instance by name"""
metric_cls = cls._metrics.get(metric_name.lower())
if metric_cls is None:
raise ValueError(f"Unknown metric: {metric_name}")
return metric_cls(**kwargs)
class MetricPresets:
"""Available metric combinations for evaluation"""
BASIC = [
"presentationmatch",
"presentationfound",
"presentationidx",
"pagematch",
"pagefound",
"presentationcount",
"bestchunkmatch",
]
LLM = ["llmrelevance"]
ALL = BASIC + LLM
@classmethod
def get_preset(cls, name: str) -> List[str]:
"""Get metric names from preset"""
try:
return getattr(cls, name.upper())
except AttributeError:
raise ValueError(f"Unknown preset name: {name}")
@classmethod
def parse_specs(cls, specs: List[str]) -> List[str]:
"""Parse metric specifications
Args:
specs: List of metric specifications. Each item can be:
- Preset name: "basic", "llm", "full"
- Metric name: "presentationmatch", "llmrelevance", etc
Returns:
List of metric names with duplicates removed
"""
metrics = []
for spec in specs:
# Check if spec is a preset name
if hasattr(cls, spec.upper()):
metrics.extend(cls.get_preset(spec))
else:
metrics.append(spec.lower())
# Remove duplicates while preserving order
seen = set()
return [m for m in metrics if not (m in seen or seen.add(m))] # type: ignore
class MlflowConfig(BaseModel):
"""Configuration for RAG evaluation"""
experiment_name: str = "RAG_test"
tracking_uri: str = f"sqlite:////{Config().navigator.eval_runs / 'mlruns.db'}"
artifacts_uri: str = f"file:////{Config().navigator.eval_artifacts}"
scorers: List[ScorerTypes]
retriever: Union[PresentationRetriever, LLMPresentationRetriever]
metrics: List[str] = ["presentationmatch", "pagematch"]
n_judge_contexts: int = 10
write_to_google: bool = False
metric_args: Dict[str, Any] = {}
model_config = ConfigDict(arbitrary_types_allowed=True)
def get_retriever_with_scorer(self, scorer: ScorerTypes) -> PresentationRetriever:
self.retriever.set_scorer(scorer)
return self.retriever
def model_post_init(self, __context: Any) -> None:
"""Process metric specifications after initialization"""
self.metrics = MetricPresets.parse_specs(self.metrics)
logger.info(f"Using metrics: {self.metrics}")
return super().model_post_init(__context)
def get_log_params(self) -> Dict[str, Any]:
"""Get parameters for MLflow logging"""
return {
"experiment_name": self.experiment_name,
"n_judge_contexts": self.n_judge_contexts,
"metrics": ",".join(self.metrics),
"metric_args": self.metric_args,
}
class RAGEvaluatorMlflow:
"""MLFlow-based evaluator for RAG pipeline"""
def __init__(
self,
config: MlflowConfig,
llm: Optional[ChatOpenAI] = None,
max_concurrent: int = 5,
):
load_dotenv()
# Setup logging
self._logger = logging.getLogger(f"{__name__}.{self.__class__.__name__}")
# Setup Evaluation
self.config = config
self.llm = llm or Config().model_config.load_vsegpt(model="openai/gpt-4o-mini")
self._max_concurrent = max_concurrent
# Setup GoogleSheets
eval_spreadsheet_id = os.getenv("EVAL_SPREADSHEET_ID")
if eval_spreadsheet_id is not None:
self.gsheets = GoogleSpreadsheetManagerMLFlow(eval_spreadsheet_id)
else:
raise FileNotFoundError("no eval_spreadsheet_id in .env")
# Setup MLFlow
mlflow.set_tracking_uri(config.tracking_uri)
mlflow.config.enable_async_logging(True)
self._logger.info(
f"MLflow tracking URI: {config.tracking_uri}, artifacts: {config.artifacts_uri}"
)
# Initialize metrics
self.metrics: List[BaseMetric] = []
for metric_name in config.metrics:
kwargs = {}
if "llm" in metric_name and llm:
kwargs = dict(llm=self.llm, n_contexts=config.n_judge_contexts)
self.metrics.append(MetricsRegistry.create(metric_name, **kwargs))
self._logger.info(f"Initialized metric: {metric_name}")
@staticmethod
def load_questions_from_sheet(*args, **kwargs) -> pd.DataFrame:
"""Load evaluation questions from spreadsheet"""
df = load_spreadsheet(*args, **kwargs)
df.fillna(dict(page=""), inplace=True)
return df
async def evaluate_single(
self, output: Dict[str, Any], question: str, ground_truth: Dict
) -> Dict[str, MetricResult]:
"""Evaluate single search result against ground truth.
Args:
output: Dictionary with retrieval results including:
- contexts: List of presentation results with metadata
question: Original search query
ground_truth: Dictionary with:
- pres_name: Expected presentation name
- pages: List of expected page numbers
- question: Original question
Returns:
Dictionary mapping metric names to MetricResult objects
"""
# Log evaluation start
self._logger.info(f"Evaluating question: '{question}'")
results = {}
# Calculate each metric
for metric in self.metrics:
try:
result = await metric.acalculate(output, ground_truth)
results[metric.name] = result
# Log metric result
log_msg = f"Metric {metric.name}: {result.score}"
if result.explanation:
log_msg += f" ({result.explanation[:200]})"
self._logger.info(log_msg)
except Exception as e:
self._logger.error(
f"Failed to calculate metric {metric.name}: {str(e)}"
)
# Create failure result
results[metric.name] = MetricResult(
name=metric.name,
score=0.0,
explanation=f"Calculation failed: {str(e)}",
)
return results
async def process_question(
self,
retriever: RetrieverTypes,
row: pd.Series,
metric_values: Dict[str, List[float]],
results_log: List[Dict],
question_idx: int,
total_questions: int,
semaphore: asyncio.Semaphore,
) -> None:
"""Process single question with semaphore-controlled concurrency"""
async with semaphore:
self._logger.info(
f"Processing question {question_idx+1}/{total_questions}: "
f"{row['question'][:50]}..."
)
ground_truth = {
"question": row["question"],
"pres_name": row["pres_name"],
"pages": [int(x) for x in row["page"].split(",") if x],
"content_type": row["content"],
}
try:
# Retrieve asynchronously
output = await retriever.aretrieve(query=row["question"])
# Evaluate results
results = await self.evaluate_single(
output=output,
question=row["question"],
ground_truth=ground_truth,
)
# Update aggregated results
result_row = {
"question": row["question"],
"expected_content_type": row["content"],
"expected_presentation": row["pres_name"],
"expected_pages": row["page"],
"retrieved_presentations": [
p["pres_name"] for p in output["contexts"]
],
"retrieved_pages": [
",".join(map(str, p["pages"])) for p in output["contexts"]
],
"best_chunk_type": output["contexts"][0]["best_chunk"]["chunk_type"], # fmt: skip
}
for metric_name, metric_result in results.items():
result_row[f"metric_{metric_name}_score"] = metric_result.score
if metric_result.explanation:
result_row[f"metric_{metric_name}_explanation"] = (
metric_result.explanation
)
metric_values[metric_name].append(metric_result.score)
results_log.append(result_row)
except Exception as e:
self._logger.error(
f"Failed to process question {question_idx+1}: {str(e)}"
)
async def process_questions_batch(
self,
retriever: RetrieverTypes,
questions_df: pd.DataFrame,
metric_values: Dict[str, List[float]],
results_log: List[Dict],
) -> None:
"""Process questions with controlled concurrency"""
# Create semaphore within the async context
semaphore = asyncio.Semaphore(self._max_concurrent)
# Create tasks for all questions
tasks = [
self.process_question(
retriever=retriever,
row=row,
metric_values=metric_values,
results_log=results_log,
question_idx=idx,
total_questions=len(questions_df),
semaphore=semaphore,
)
for idx, (_, row) in enumerate(questions_df.iterrows())
]
# Wait for all tasks to complete
await tqdm_asyncio.gather(
*tasks,
desc=f"Processing questions for '{retriever.scorer.id[:15]}' (max {self._max_concurrent} concurrent)",
total=len(questions_df),
dynamic_ncols=True, # Adjust width automatically
)
def run_evaluation(self, questions_df: pd.DataFrame) -> None:
"""Run evaluation with async LLM queries and controlled concurrency"""
timestamp = datetime.now().replace(microsecond=0).isoformat()
self._logger.info(f"Starting evaluation with {len(questions_df)} questions")
# MLflow setup
experiment = mlflow.get_experiment_by_name(self.config.experiment_name)
if experiment is not None:
experiment_id = experiment.experiment_id
self._logger.info(
f"Using existing experiment: {self.config.experiment_name}"
)
else:
experiment_id = mlflow.create_experiment(
self.config.experiment_name,
artifact_location=self.config.artifacts_uri,
)
experiment = mlflow.get_experiment_by_name(self.config.experiment_name)
self._logger.info(f"Created new experiment: {self.config.experiment_name}")
mlflow.set_experiment(experiment_id=experiment_id)
for scorer in self.config.scorers:
self._logger.info(f"Evaluating with scorer: {scorer.id}")
# Initialize retriever
retriever = self.config.get_retriever_with_scorer(scorer)
# Get preprocessor id
preprocessor_id = (
retriever.storage.query_preprocessor.id
if retriever.storage.query_preprocessor
else "None"
)
with mlflow.start_run(
run_name=f"scorer_{scorer.id}__retriever_{retriever.id}__preprocessor_{preprocessor_id}"
):
# Log preprocessor
mlflow.log_params({"preprocessing": preprocessor_id})
self._logger.info(f"Using preprocessor: {preprocessor_id}")
# Log config parameters
mlflow.log_params(
{f"config_{k}": v for k, v in self.config.get_log_params().items()}
)
self._logger.debug("Logged config parameters")
# Log scorer parameters
mlflow.log_params(
{f"scorer_{k}": v for k, v in scorer.model_dump().items()}
)
self._logger.debug("Logged scorer parameters")
# Initialize retriever and log its parameters
mlflow.log_params(
{f"retriever_{k}": v for k, v in retriever.get_log_params().items()}
)
self._logger.debug("Logged retriever parameters")
# Initialize aggregation containers
results_log = []
metric_values = {m.name: [] for m in self.metrics}
# Process questions with async handling
asyncio.run(
self.process_questions_batch(
retriever, questions_df, metric_values, results_log
)
)
# Calculate n_errors
n_errors = (
len(questions_df) - len(results_log)
if results_log
else len(questions_df)
)
if n_errors > 1:
logger.error(f"{n_errors} while processing {retriever.id}")
# Process results
results_df = pd.DataFrame(results_log)
results_df["experiment_name"] = (
experiment.name if experiment is not None else "no_name"
)
results_df["experiment_id"] = experiment_id
results_df["scorer"] = scorer.id
results_df["retriever"] = retriever.id
results_df["timestamp"] = timestamp
# Save results
with NamedTemporaryFile(mode="w", suffix=".csv", delete=False) as f:
results_df.to_csv(f.name, index=False)
fpath = str("detailed_results")
mlflow.log_artifact(f.name, fpath)
self._logger.info(f"Saved detailed results to {fpath}")
# Write to google sheets if enabled
if self.config.write_to_google:
self.gsheets.write_evaluation_results(
results_df=results_df,
metric_values=metric_values,
experiment_name=self.config.experiment_name,
)
# Log metrics
for name, values in metric_values.items():
if values:
mean_value = np.nanmean(values)
mlflow.log_metric(f"mean_{name}", mean_value)
mlflow.log_metric(f"n_questions", len(questions_df))
mlflow.log_metric(f"error_rate", n_errors / len(questions_df))
self._logger.info(f"Mean {name}: {mean_value:.3f}")