Spaces:
Sleeping
fix(pipeline/metrics): 3 bugs pipelines OCR+LLM
Browse filesBug 2 — CER 0.00% pour hypothèse vide (root cause)
metrics.py : ajout d'une garde explicite avant les appels jiwer.
Si l'hypothèse est vide (ou uniquement des espaces) avec une référence
non vide, retourne immédiatement CER=WER=MER=WIL=1.0.
jiwer.wer("ref", "") lève ZeroDivisionError ; l'ancien except retournait
silencieusement cer=0.0 au lieu de 1.0.
Bug 1 — Sortie LLM vide sans avertissement
mistral_adapter.py :
- WARNING logu si response.choices[0].message.content est vide ou None
- DEBUG logs : longueur prompt, longueur réponse, extrait réponse
- Constante _TEXT_ONLY_MODELS pour ministral-3b/8b-latest et équivalents
- WARNING + neutralisation de l'image si modèle text-only reçoit image_b64
pipelines/base.py :
- WARNING si le LLM retourne un texte vide (avec conseils de diagnostic)
- WARNING si le moteur OCR produit un texte vide avant l'envoi au LLM
- DEBUG logs : longueur texte OCR, longueur réponse LLM, extrait
Bug 3 — Divergence runner/rapport
Résolu par le fix Bug 2 : le DocumentResult stocke désormais CER=1.0
(via compute_metrics) pour toute hypothèse vide, garantissant la
cohérence entre le log runner et l'affichage dans le rapport HTML.
Tests : +20 tests (test_metrics.py +4, test_sprint15_llm_pipeline_bugs.py +16)
890 tests passent.
https://claude.ai/code/session_017gXea9mxBQqDTAsSQd7aAq
- picarones/core/metrics.py +12 -0
- picarones/llm/mistral_adapter.py +56 -1
- picarones/pipelines/base.py +43 -1
- tests/test_metrics.py +28 -0
- tests/test_sprint15_llm_pipeline_bugs.py +291 -0
|
@@ -152,6 +152,18 @@ def compute_metrics(
|
|
| 152 |
error="jiwer n'est pas installé (pip install jiwer)",
|
| 153 |
)
|
| 154 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 155 |
try:
|
| 156 |
# Exclusion de caractères avant tout calcul
|
| 157 |
if char_exclude:
|
|
|
|
| 152 |
error="jiwer n'est pas installé (pip install jiwer)",
|
| 153 |
)
|
| 154 |
|
| 155 |
+
# Hypothèse vide avec référence non vide = erreur totale (toutes les
|
| 156 |
+
# métriques jiwer lèvent une ZeroDivisionError sur hypothèse vide).
|
| 157 |
+
ref_stripped = reference.strip()
|
| 158 |
+
hyp_stripped = hypothesis.strip() if hypothesis else ""
|
| 159 |
+
if ref_stripped and not hyp_stripped:
|
| 160 |
+
return MetricsResult(
|
| 161 |
+
cer=1.0, cer_nfc=1.0, cer_caseless=1.0,
|
| 162 |
+
wer=1.0, wer_normalized=1.0, mer=1.0, wil=1.0,
|
| 163 |
+
reference_length=len(reference),
|
| 164 |
+
hypothesis_length=0,
|
| 165 |
+
)
|
| 166 |
+
|
| 167 |
try:
|
| 168 |
# Exclusion de caractères avant tout calcul
|
| 169 |
if char_exclude:
|
|
@@ -2,11 +2,25 @@
|
|
| 2 |
|
| 3 |
from __future__ import annotations
|
| 4 |
|
|
|
|
| 5 |
import os
|
| 6 |
from typing import Optional
|
| 7 |
|
| 8 |
from picarones.llm.base import BaseLLMAdapter
|
| 9 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 10 |
|
| 11 |
class MistralAdapter(BaseLLMAdapter):
|
| 12 |
"""Adaptateur pour les modèles Mistral AI.
|
|
@@ -15,6 +29,11 @@ class MistralAdapter(BaseLLMAdapter):
|
|
| 15 |
|
| 16 |
Modes supportés : text_only (tous modèles), text_and_image et zero_shot
|
| 17 |
avec les modèles multimodaux (pixtral-12b, pixtral-large).
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 18 |
"""
|
| 19 |
|
| 20 |
@property
|
|
@@ -32,6 +51,11 @@ class MistralAdapter(BaseLLMAdapter):
|
|
| 32 |
) -> None:
|
| 33 |
super().__init__(model, config)
|
| 34 |
self._api_key = os.environ.get("MISTRAL_API_KEY")
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 35 |
|
| 36 |
def _call(self, prompt: str, image_b64: Optional[str] = None) -> str:
|
| 37 |
if not self._api_key:
|
|
@@ -49,6 +73,15 @@ class MistralAdapter(BaseLLMAdapter):
|
|
| 49 |
temperature = float(self.config.get("temperature", 0.0))
|
| 50 |
max_tokens = int(self.config.get("max_tokens", 4096))
|
| 51 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 52 |
if image_b64:
|
| 53 |
content: list | str = [
|
| 54 |
{"type": "text", "text": prompt},
|
|
@@ -60,10 +93,32 @@ class MistralAdapter(BaseLLMAdapter):
|
|
| 60 |
else:
|
| 61 |
content = prompt
|
| 62 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 63 |
response = client.chat.complete(
|
| 64 |
model=self.model,
|
| 65 |
messages=[{"role": "user", "content": content}],
|
| 66 |
temperature=temperature,
|
| 67 |
max_tokens=max_tokens,
|
| 68 |
)
|
| 69 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 2 |
|
| 3 |
from __future__ import annotations
|
| 4 |
|
| 5 |
+
import logging
|
| 6 |
import os
|
| 7 |
from typing import Optional
|
| 8 |
|
| 9 |
from picarones.llm.base import BaseLLMAdapter
|
| 10 |
|
| 11 |
+
logger = logging.getLogger(__name__)
|
| 12 |
+
|
| 13 |
+
# Modèles Mistral qui NE supportent PAS l'API chat/completions multimodale.
|
| 14 |
+
# Ces petits modèles sont text-only; le passer avec une image provoque une erreur.
|
| 15 |
+
_TEXT_ONLY_MODELS = frozenset({
|
| 16 |
+
"ministral-3b-latest",
|
| 17 |
+
"ministral-8b-latest",
|
| 18 |
+
"mistral-tiny",
|
| 19 |
+
"mistral-tiny-latest",
|
| 20 |
+
"open-mistral-7b",
|
| 21 |
+
"open-mixtral-8x7b",
|
| 22 |
+
})
|
| 23 |
+
|
| 24 |
|
| 25 |
class MistralAdapter(BaseLLMAdapter):
|
| 26 |
"""Adaptateur pour les modèles Mistral AI.
|
|
|
|
| 29 |
|
| 30 |
Modes supportés : text_only (tous modèles), text_and_image et zero_shot
|
| 31 |
avec les modèles multimodaux (pixtral-12b, pixtral-large).
|
| 32 |
+
|
| 33 |
+
Note
|
| 34 |
+
----
|
| 35 |
+
Les modèles ``ministral-3b-latest`` et ``ministral-8b-latest`` ne supportent
|
| 36 |
+
pas le mode multimodal — utiliser ``PipelineMode.TEXT_ONLY`` avec ces modèles.
|
| 37 |
"""
|
| 38 |
|
| 39 |
@property
|
|
|
|
| 51 |
) -> None:
|
| 52 |
super().__init__(model, config)
|
| 53 |
self._api_key = os.environ.get("MISTRAL_API_KEY")
|
| 54 |
+
if self.model in _TEXT_ONLY_MODELS:
|
| 55 |
+
logger.info(
|
| 56 |
+
"[MistralAdapter] modèle '%s' : text-only (pas de support multimodal).",
|
| 57 |
+
self.model,
|
| 58 |
+
)
|
| 59 |
|
| 60 |
def _call(self, prompt: str, image_b64: Optional[str] = None) -> str:
|
| 61 |
if not self._api_key:
|
|
|
|
| 73 |
temperature = float(self.config.get("temperature", 0.0))
|
| 74 |
max_tokens = int(self.config.get("max_tokens", 4096))
|
| 75 |
|
| 76 |
+
# Les modèles text-only ne supportent pas les images
|
| 77 |
+
if image_b64 and self.model in _TEXT_ONLY_MODELS:
|
| 78 |
+
logger.warning(
|
| 79 |
+
"[MistralAdapter] modèle '%s' ne supporte pas les images — "
|
| 80 |
+
"image ignorée, appel en mode texte seul.",
|
| 81 |
+
self.model,
|
| 82 |
+
)
|
| 83 |
+
image_b64 = None
|
| 84 |
+
|
| 85 |
if image_b64:
|
| 86 |
content: list | str = [
|
| 87 |
{"type": "text", "text": prompt},
|
|
|
|
| 93 |
else:
|
| 94 |
content = prompt
|
| 95 |
|
| 96 |
+
logger.debug(
|
| 97 |
+
"[MistralAdapter] appel %s — longueur prompt : %d caractères, image : %s",
|
| 98 |
+
self.model, len(prompt), "oui" if image_b64 else "non",
|
| 99 |
+
)
|
| 100 |
+
|
| 101 |
response = client.chat.complete(
|
| 102 |
model=self.model,
|
| 103 |
messages=[{"role": "user", "content": content}],
|
| 104 |
temperature=temperature,
|
| 105 |
max_tokens=max_tokens,
|
| 106 |
)
|
| 107 |
+
raw = response.choices[0].message.content
|
| 108 |
+
text = raw or ""
|
| 109 |
+
|
| 110 |
+
if not text or not text.strip():
|
| 111 |
+
logger.warning(
|
| 112 |
+
"[MistralAdapter] réponse vide reçue du modèle '%s' "
|
| 113 |
+
"(longueur brute : %s). "
|
| 114 |
+
"Vérifier que le modèle supporte l'API chat/completions et "
|
| 115 |
+
"que le prompt contient bien {ocr_output}.",
|
| 116 |
+
self.model, len(raw) if raw is not None else "None",
|
| 117 |
+
)
|
| 118 |
+
else:
|
| 119 |
+
logger.debug(
|
| 120 |
+
"[MistralAdapter] réponse reçue — %d caractères, extrait : %r",
|
| 121 |
+
len(text), text[:120],
|
| 122 |
+
)
|
| 123 |
+
|
| 124 |
+
return text
|
|
@@ -15,6 +15,7 @@ exposées via ``EngineResult.metadata``.
|
|
| 15 |
from __future__ import annotations
|
| 16 |
|
| 17 |
import base64
|
|
|
|
| 18 |
import time
|
| 19 |
from enum import Enum
|
| 20 |
from pathlib import Path
|
|
@@ -23,6 +24,8 @@ from typing import Optional
|
|
| 23 |
from picarones.engines.base import BaseOCREngine, EngineResult
|
| 24 |
from picarones.llm.base import BaseLLMAdapter
|
| 25 |
|
|
|
|
|
|
|
| 26 |
|
| 27 |
class PipelineMode(str, Enum):
|
| 28 |
"""Mode d'appel LLM dans le pipeline."""
|
|
@@ -146,6 +149,9 @@ class OCRLLMPipeline(BaseOCREngine):
|
|
| 146 |
if self.mode == PipelineMode.ZERO_SHOT:
|
| 147 |
image_b64 = _image_to_b64(image_path)
|
| 148 |
prompt = self._build_prompt(image_b64=image_b64)
|
|
|
|
|
|
|
|
|
|
| 149 |
result = self.llm_adapter.complete(prompt, image_b64=image_b64)
|
| 150 |
|
| 151 |
elif self.mode == PipelineMode.TEXT_ONLY:
|
|
@@ -154,6 +160,16 @@ class OCRLLMPipeline(BaseOCREngine):
|
|
| 154 |
ocr_result = self.ocr_engine.run(image_path)
|
| 155 |
ocr_text = ocr_result.text
|
| 156 |
self._last_ocr_text = ocr_text
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 157 |
prompt = self._build_prompt(ocr_text=ocr_text)
|
| 158 |
result = self.llm_adapter.complete(prompt)
|
| 159 |
|
|
@@ -163,6 +179,16 @@ class OCRLLMPipeline(BaseOCREngine):
|
|
| 163 |
ocr_result = self.ocr_engine.run(image_path)
|
| 164 |
ocr_text = ocr_result.text
|
| 165 |
self._last_ocr_text = ocr_text
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 166 |
image_b64 = _image_to_b64(image_path)
|
| 167 |
prompt = self._build_prompt(ocr_text=ocr_text, image_b64=image_b64)
|
| 168 |
result = self.llm_adapter.complete(prompt, image_b64=image_b64)
|
|
@@ -170,7 +196,23 @@ class OCRLLMPipeline(BaseOCREngine):
|
|
| 170 |
if not result.success:
|
| 171 |
raise RuntimeError(f"Erreur LLM ({self.llm_adapter.model}): {result.error}")
|
| 172 |
|
| 173 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 174 |
|
| 175 |
# ------------------------------------------------------------------
|
| 176 |
# Override run() pour injecter les métadonnées pipeline
|
|
|
|
| 15 |
from __future__ import annotations
|
| 16 |
|
| 17 |
import base64
|
| 18 |
+
import logging
|
| 19 |
import time
|
| 20 |
from enum import Enum
|
| 21 |
from pathlib import Path
|
|
|
|
| 24 |
from picarones.engines.base import BaseOCREngine, EngineResult
|
| 25 |
from picarones.llm.base import BaseLLMAdapter
|
| 26 |
|
| 27 |
+
logger = logging.getLogger(__name__)
|
| 28 |
+
|
| 29 |
|
| 30 |
class PipelineMode(str, Enum):
|
| 31 |
"""Mode d'appel LLM dans le pipeline."""
|
|
|
|
| 149 |
if self.mode == PipelineMode.ZERO_SHOT:
|
| 150 |
image_b64 = _image_to_b64(image_path)
|
| 151 |
prompt = self._build_prompt(image_b64=image_b64)
|
| 152 |
+
logger.debug(
|
| 153 |
+
"[%s] zero-shot — longueur prompt : %d car.", self._name, len(prompt)
|
| 154 |
+
)
|
| 155 |
result = self.llm_adapter.complete(prompt, image_b64=image_b64)
|
| 156 |
|
| 157 |
elif self.mode == PipelineMode.TEXT_ONLY:
|
|
|
|
| 160 |
ocr_result = self.ocr_engine.run(image_path)
|
| 161 |
ocr_text = ocr_result.text
|
| 162 |
self._last_ocr_text = ocr_text
|
| 163 |
+
logger.debug(
|
| 164 |
+
"[%s] texte OCR : %d car. → envoi au LLM.",
|
| 165 |
+
self._name, len(ocr_text),
|
| 166 |
+
)
|
| 167 |
+
if not ocr_text.strip():
|
| 168 |
+
logger.warning(
|
| 169 |
+
"[%s] le moteur OCR a produit un texte vide pour '%s'. "
|
| 170 |
+
"Le LLM recevra un prompt sans texte OCR ({ocr_output} vide).",
|
| 171 |
+
self._name, image_path.name,
|
| 172 |
+
)
|
| 173 |
prompt = self._build_prompt(ocr_text=ocr_text)
|
| 174 |
result = self.llm_adapter.complete(prompt)
|
| 175 |
|
|
|
|
| 179 |
ocr_result = self.ocr_engine.run(image_path)
|
| 180 |
ocr_text = ocr_result.text
|
| 181 |
self._last_ocr_text = ocr_text
|
| 182 |
+
logger.debug(
|
| 183 |
+
"[%s] texte OCR : %d car. + image → envoi au LLM.",
|
| 184 |
+
self._name, len(ocr_text),
|
| 185 |
+
)
|
| 186 |
+
if not ocr_text.strip():
|
| 187 |
+
logger.warning(
|
| 188 |
+
"[%s] le moteur OCR a produit un texte vide pour '%s'. "
|
| 189 |
+
"Le LLM recevra un prompt sans texte OCR ({ocr_output} vide).",
|
| 190 |
+
self._name, image_path.name,
|
| 191 |
+
)
|
| 192 |
image_b64 = _image_to_b64(image_path)
|
| 193 |
prompt = self._build_prompt(ocr_text=ocr_text, image_b64=image_b64)
|
| 194 |
result = self.llm_adapter.complete(prompt, image_b64=image_b64)
|
|
|
|
| 196 |
if not result.success:
|
| 197 |
raise RuntimeError(f"Erreur LLM ({self.llm_adapter.model}): {result.error}")
|
| 198 |
|
| 199 |
+
llm_text = result.text
|
| 200 |
+
if not llm_text or not llm_text.strip():
|
| 201 |
+
logger.warning(
|
| 202 |
+
"[%s] le LLM ('%s') a retourné un texte vide pour '%s'. "
|
| 203 |
+
"CER sera calculé à 1.0 (100%%). "
|
| 204 |
+
"Vérifier : (1) le prompt contient-il {ocr_output} ? "
|
| 205 |
+
"(2) le modèle supporte-t-il ce mode d'appel ? "
|
| 206 |
+
"(3) la réponse n'est-elle pas tronquée (max_tokens) ?",
|
| 207 |
+
self._name, self.llm_adapter.model, image_path.name,
|
| 208 |
+
)
|
| 209 |
+
else:
|
| 210 |
+
logger.debug(
|
| 211 |
+
"[%s] réponse LLM : %d car., extrait : %r",
|
| 212 |
+
self._name, len(llm_text), llm_text[:120],
|
| 213 |
+
)
|
| 214 |
+
|
| 215 |
+
return llm_text
|
| 216 |
|
| 217 |
# ------------------------------------------------------------------
|
| 218 |
# Override run() pour injecter les métadonnées pipeline
|
|
@@ -74,6 +74,34 @@ class TestComputeMetrics:
|
|
| 74 |
result = compute_metrics("abcd", "abce")
|
| 75 |
assert result.cer_percent == pytest.approx(25.0, rel=1e-2)
|
| 76 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 77 |
|
| 78 |
class TestAggregateMetrics:
|
| 79 |
"""Tests de aggregate_metrics."""
|
|
|
|
| 74 |
result = compute_metrics("abcd", "abce")
|
| 75 |
assert result.cer_percent == pytest.approx(25.0, rel=1e-2)
|
| 76 |
|
| 77 |
+
# ── Bug fix : hypothèse vide → CER doit être 1.0, pas 0.0 (bug sprint 13) ──
|
| 78 |
+
|
| 79 |
+
def test_empty_hypothesis_cer_is_one(self):
|
| 80 |
+
"""Hypothèse vide avec référence non vide doit donner CER=1.0."""
|
| 81 |
+
result = compute_metrics("Bonjour le monde", "")
|
| 82 |
+
assert result.cer == pytest.approx(1.0), (
|
| 83 |
+
f"CER attendu 1.0 pour hypothèse vide, obtenu {result.cer}"
|
| 84 |
+
)
|
| 85 |
+
assert result.error is None
|
| 86 |
+
|
| 87 |
+
def test_empty_hypothesis_wer_is_one(self):
|
| 88 |
+
"""WER doit être 1.0 pour hypothèse vide (pas de ZeroDivisionError)."""
|
| 89 |
+
result = compute_metrics("hello world", "")
|
| 90 |
+
assert result.wer == pytest.approx(1.0)
|
| 91 |
+
assert result.mer == pytest.approx(1.0)
|
| 92 |
+
assert result.wil == pytest.approx(1.0)
|
| 93 |
+
assert result.error is None
|
| 94 |
+
|
| 95 |
+
def test_empty_hypothesis_whitespace_is_treated_as_empty(self):
|
| 96 |
+
"""Hypothèse avec uniquement des espaces est traitée comme vide."""
|
| 97 |
+
result = compute_metrics("Bonjour", " ")
|
| 98 |
+
assert result.cer == pytest.approx(1.0)
|
| 99 |
+
|
| 100 |
+
def test_empty_hypothesis_hypothesis_length_is_zero(self):
|
| 101 |
+
"""hypothesis_length doit être 0 pour hypothèse vide."""
|
| 102 |
+
result = compute_metrics("Bonjour le monde", "")
|
| 103 |
+
assert result.hypothesis_length == 0
|
| 104 |
+
|
| 105 |
|
| 106 |
class TestAggregateMetrics:
|
| 107 |
"""Tests de aggregate_metrics."""
|
|
@@ -0,0 +1,291 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""Tests pour le sprint 15 — Correction des bugs dans les pipelines OCR+LLM.
|
| 2 |
+
|
| 3 |
+
Bug 1 : Sortie LLM vide → WARNING logué + pas de crash
|
| 4 |
+
Bug 2 : CER 0.00% pour hypothèse vide → doit être 1.0 (100%)
|
| 5 |
+
Bug 3 : Divergence runner/rapport → cohérence des métriques
|
| 6 |
+
"""
|
| 7 |
+
from __future__ import annotations
|
| 8 |
+
|
| 9 |
+
import logging
|
| 10 |
+
from pathlib import Path
|
| 11 |
+
from unittest.mock import MagicMock, patch
|
| 12 |
+
|
| 13 |
+
import pytest
|
| 14 |
+
|
| 15 |
+
|
| 16 |
+
# ---------------------------------------------------------------------------
|
| 17 |
+
# Bug 2 — compute_metrics : hypothèse vide
|
| 18 |
+
# ---------------------------------------------------------------------------
|
| 19 |
+
|
| 20 |
+
class TestEmptyHypothesisMetrics:
|
| 21 |
+
"""compute_metrics doit retourner CER=1.0, pas 0.0, pour hypothèse vide."""
|
| 22 |
+
|
| 23 |
+
def test_empty_hypothesis_cer_is_one(self):
|
| 24 |
+
from picarones.core.metrics import compute_metrics
|
| 25 |
+
result = compute_metrics("Bonjour le monde", "")
|
| 26 |
+
assert result.cer == pytest.approx(1.0)
|
| 27 |
+
assert result.error is None
|
| 28 |
+
|
| 29 |
+
def test_empty_hypothesis_all_metrics_are_one(self):
|
| 30 |
+
from picarones.core.metrics import compute_metrics
|
| 31 |
+
result = compute_metrics("hello world", "")
|
| 32 |
+
assert result.cer == pytest.approx(1.0)
|
| 33 |
+
assert result.wer == pytest.approx(1.0)
|
| 34 |
+
assert result.mer == pytest.approx(1.0)
|
| 35 |
+
assert result.wil == pytest.approx(1.0)
|
| 36 |
+
|
| 37 |
+
def test_whitespace_only_hypothesis_cer_is_one(self):
|
| 38 |
+
from picarones.core.metrics import compute_metrics
|
| 39 |
+
result = compute_metrics("Bonjour", " \t\n")
|
| 40 |
+
assert result.cer == pytest.approx(1.0)
|
| 41 |
+
|
| 42 |
+
def test_none_hypothesis_guarded(self):
|
| 43 |
+
"""compute_metrics ne doit pas planter si hypothesis=None."""
|
| 44 |
+
from picarones.core.metrics import compute_metrics
|
| 45 |
+
# None ne sera jamais passé en pratique, mais on teste la robustesse
|
| 46 |
+
# via une chaîne vide (le runner convertit None → "")
|
| 47 |
+
result = compute_metrics("test", "")
|
| 48 |
+
assert result.cer == pytest.approx(1.0)
|
| 49 |
+
|
| 50 |
+
def test_both_empty_cer_is_zero(self):
|
| 51 |
+
"""Référence ET hypothèse vides → CER=0.0 (pas d'erreur à mesurer)."""
|
| 52 |
+
from picarones.core.metrics import compute_metrics
|
| 53 |
+
result = compute_metrics("", "")
|
| 54 |
+
assert result.cer == pytest.approx(0.0)
|
| 55 |
+
|
| 56 |
+
def test_empty_reference_nonempty_hypothesis(self):
|
| 57 |
+
"""Référence vide avec hypothèse non vide → CER=1.0 (comportement existant)."""
|
| 58 |
+
from picarones.core.metrics import compute_metrics
|
| 59 |
+
result = compute_metrics("", "something")
|
| 60 |
+
assert result.cer == pytest.approx(1.0)
|
| 61 |
+
|
| 62 |
+
def test_normal_case_unchanged(self):
|
| 63 |
+
"""Un cas normal ne doit pas être affecté par le guard."""
|
| 64 |
+
from picarones.core.metrics import compute_metrics
|
| 65 |
+
result = compute_metrics("abcd", "abce")
|
| 66 |
+
assert result.cer == pytest.approx(0.25)
|
| 67 |
+
assert result.error is None
|
| 68 |
+
|
| 69 |
+
|
| 70 |
+
# ---------------------------------------------------------------------------
|
| 71 |
+
# Bug 1 — MistralAdapter : WARNING pour réponse vide
|
| 72 |
+
# ---------------------------------------------------------------------------
|
| 73 |
+
|
| 74 |
+
class TestMistralAdapterLogging:
|
| 75 |
+
"""MistralAdapter doit loguer un WARNING si la réponse LLM est vide."""
|
| 76 |
+
|
| 77 |
+
def _make_mock_mistral_module(self, content: str | None):
|
| 78 |
+
"""Retourne un module mistralai simulé avec la réponse donnée."""
|
| 79 |
+
mock_response = MagicMock()
|
| 80 |
+
mock_response.choices = [MagicMock()]
|
| 81 |
+
mock_response.choices[0].message.content = content
|
| 82 |
+
|
| 83 |
+
mock_client = MagicMock()
|
| 84 |
+
mock_client.chat.complete.return_value = mock_response
|
| 85 |
+
|
| 86 |
+
MockMistralClass = MagicMock(return_value=mock_client)
|
| 87 |
+
|
| 88 |
+
import types
|
| 89 |
+
fake_module = types.ModuleType("mistralai")
|
| 90 |
+
fake_module.Mistral = MockMistralClass
|
| 91 |
+
return fake_module, mock_client
|
| 92 |
+
|
| 93 |
+
def _run_adapter(self, adapter, fake_mod, prompt="test prompt", image_b64=None):
|
| 94 |
+
"""Exécute l'adapter avec le module mistralai simulé."""
|
| 95 |
+
import sys
|
| 96 |
+
with patch.dict(sys.modules, {"mistralai": fake_mod}):
|
| 97 |
+
adapter._api_key = "fake-key" # injecter la clé directement
|
| 98 |
+
return adapter.complete(prompt, image_b64=image_b64)
|
| 99 |
+
|
| 100 |
+
def test_warning_on_empty_response(self, caplog):
|
| 101 |
+
"""Un WARNING doit être émis si le LLM retourne une chaîne vide."""
|
| 102 |
+
from picarones.llm.mistral_adapter import MistralAdapter
|
| 103 |
+
|
| 104 |
+
fake_mod, _ = self._make_mock_mistral_module("")
|
| 105 |
+
adapter = MistralAdapter(model="ministral-3b-latest")
|
| 106 |
+
|
| 107 |
+
with caplog.at_level(logging.WARNING, logger="picarones.llm.mistral_adapter"):
|
| 108 |
+
result = self._run_adapter(adapter, fake_mod)
|
| 109 |
+
|
| 110 |
+
assert result.text == ""
|
| 111 |
+
assert any(
|
| 112 |
+
"vide" in rec.message.lower() or "empty" in rec.message.lower()
|
| 113 |
+
for rec in caplog.records
|
| 114 |
+
if rec.levelno >= logging.WARNING
|
| 115 |
+
), f"WARNING attendu, messages : {[r.message for r in caplog.records]}"
|
| 116 |
+
|
| 117 |
+
def test_no_warning_on_normal_response(self, caplog):
|
| 118 |
+
"""Aucun WARNING ne doit être émis pour une réponse normale."""
|
| 119 |
+
from picarones.llm.mistral_adapter import MistralAdapter
|
| 120 |
+
|
| 121 |
+
fake_mod, _ = self._make_mock_mistral_module("Texte OCR corrigé")
|
| 122 |
+
adapter = MistralAdapter(model="ministral-3b-latest")
|
| 123 |
+
|
| 124 |
+
with caplog.at_level(logging.WARNING, logger="picarones.llm.mistral_adapter"):
|
| 125 |
+
result = self._run_adapter(adapter, fake_mod)
|
| 126 |
+
|
| 127 |
+
assert result.text == "Texte OCR corrigé"
|
| 128 |
+
assert not any(rec.levelno >= logging.WARNING for rec in caplog.records)
|
| 129 |
+
|
| 130 |
+
def test_warning_on_none_response_content(self, caplog):
|
| 131 |
+
"""WARNING doit être émis si message.content est None."""
|
| 132 |
+
from picarones.llm.mistral_adapter import MistralAdapter
|
| 133 |
+
|
| 134 |
+
fake_mod, _ = self._make_mock_mistral_module(None)
|
| 135 |
+
adapter = MistralAdapter(model="ministral-3b-latest")
|
| 136 |
+
|
| 137 |
+
with caplog.at_level(logging.WARNING, logger="picarones.llm.mistral_adapter"):
|
| 138 |
+
result = self._run_adapter(adapter, fake_mod)
|
| 139 |
+
|
| 140 |
+
assert result.text == ""
|
| 141 |
+
assert any(rec.levelno >= logging.WARNING for rec in caplog.records)
|
| 142 |
+
|
| 143 |
+
def test_text_only_models_set_exists(self):
|
| 144 |
+
"""La liste des modèles text-only doit contenir ministral-3b."""
|
| 145 |
+
from picarones.llm.mistral_adapter import _TEXT_ONLY_MODELS
|
| 146 |
+
assert "ministral-3b-latest" in _TEXT_ONLY_MODELS
|
| 147 |
+
|
| 148 |
+
def test_image_ignored_for_text_only_model(self, caplog):
|
| 149 |
+
"""L'image doit être ignorée (avec WARNING) pour un modèle text-only."""
|
| 150 |
+
from picarones.llm.mistral_adapter import MistralAdapter
|
| 151 |
+
|
| 152 |
+
fake_mod, mock_client = self._make_mock_mistral_module("résultat")
|
| 153 |
+
adapter = MistralAdapter(model="ministral-3b-latest")
|
| 154 |
+
|
| 155 |
+
with caplog.at_level(logging.WARNING, logger="picarones.llm.mistral_adapter"):
|
| 156 |
+
result = self._run_adapter(adapter, fake_mod, image_b64="fake_b64")
|
| 157 |
+
|
| 158 |
+
# L'appel doit avoir été fait SANS image (modèle text-only)
|
| 159 |
+
call_kwargs = mock_client.chat.complete.call_args
|
| 160 |
+
_, kwargs = call_kwargs
|
| 161 |
+
msg_content = kwargs.get("messages", [{}])[0].get("content", "")
|
| 162 |
+
assert isinstance(msg_content, str), "Image aurait dû être ignorée (content doit être str)"
|
| 163 |
+
# Au moins un WARNING doit mentionner l'image ignorée
|
| 164 |
+
assert any("ignor" in rec.message.lower() for rec in caplog.records
|
| 165 |
+
if rec.levelno >= logging.WARNING)
|
| 166 |
+
|
| 167 |
+
|
| 168 |
+
# ---------------------------------------------------------------------------
|
| 169 |
+
# Bug 1 — OCRLLMPipeline : WARNING quand le LLM retourne texte vide
|
| 170 |
+
# ---------------------------------------------------------------------------
|
| 171 |
+
|
| 172 |
+
class TestPipelineEmptyLLMResponse:
|
| 173 |
+
"""Le pipeline doit loguer un WARNING si le LLM retourne un texte vide."""
|
| 174 |
+
|
| 175 |
+
def _make_pipeline(self, llm_text: str):
|
| 176 |
+
"""Crée un pipeline dont le LLM retourne llm_text."""
|
| 177 |
+
from picarones.pipelines.base import OCRLLMPipeline, PipelineMode
|
| 178 |
+
from picarones.llm.base import LLMResult
|
| 179 |
+
|
| 180 |
+
mock_ocr = MagicMock()
|
| 181 |
+
mock_ocr.name = "mock_ocr"
|
| 182 |
+
mock_ocr.run.return_value = MagicMock(text="texte ocr brut", error=None, success=True)
|
| 183 |
+
mock_ocr._safe_version.return_value = "1.0"
|
| 184 |
+
|
| 185 |
+
mock_llm = MagicMock()
|
| 186 |
+
mock_llm.name = "mock_llm"
|
| 187 |
+
mock_llm.model = "mock-model"
|
| 188 |
+
mock_llm.complete.return_value = LLMResult(
|
| 189 |
+
model_id="mock-model", text=llm_text, duration_seconds=0.1,
|
| 190 |
+
)
|
| 191 |
+
|
| 192 |
+
return OCRLLMPipeline(
|
| 193 |
+
ocr_engine=mock_ocr,
|
| 194 |
+
llm_adapter=mock_llm,
|
| 195 |
+
mode=PipelineMode.TEXT_ONLY,
|
| 196 |
+
prompt="correction_medieval_french.txt",
|
| 197 |
+
)
|
| 198 |
+
|
| 199 |
+
def test_warning_on_empty_llm_output(self, tmp_path, caplog):
|
| 200 |
+
"""WARNING doit être logu si le LLM retourne une chaîne vide."""
|
| 201 |
+
import shutil
|
| 202 |
+
# Créer une fausse image
|
| 203 |
+
img_path = tmp_path / "test.png"
|
| 204 |
+
img_path.write_bytes(b"\x89PNG\r\n\x1a\n" + b"\x00" * 100)
|
| 205 |
+
|
| 206 |
+
pipeline = self._make_pipeline("")
|
| 207 |
+
with caplog.at_level(logging.WARNING, logger="picarones.pipelines.base"):
|
| 208 |
+
result = pipeline.run(img_path)
|
| 209 |
+
|
| 210 |
+
assert result.text == ""
|
| 211 |
+
assert any(
|
| 212 |
+
"vide" in rec.message.lower() or "empty" in rec.message.lower()
|
| 213 |
+
for rec in caplog.records
|
| 214 |
+
if rec.levelno >= logging.WARNING
|
| 215 |
+
), f"WARNING attendu, messages : {[r.message for r in caplog.records]}"
|
| 216 |
+
|
| 217 |
+
def test_no_warning_on_normal_llm_output(self, tmp_path, caplog):
|
| 218 |
+
"""Aucun WARNING ne doit être émis pour une sortie LLM normale."""
|
| 219 |
+
img_path = tmp_path / "test.png"
|
| 220 |
+
img_path.write_bytes(b"\x89PNG\r\n\x1a\n" + b"\x00" * 100)
|
| 221 |
+
|
| 222 |
+
pipeline = self._make_pipeline("Texte corrigé par le LLM")
|
| 223 |
+
with caplog.at_level(logging.WARNING, logger="picarones.pipelines.base"):
|
| 224 |
+
result = pipeline.run(img_path)
|
| 225 |
+
|
| 226 |
+
assert result.text == "Texte corrigé par le LLM"
|
| 227 |
+
assert not any(
|
| 228 |
+
"vide" in rec.message.lower()
|
| 229 |
+
for rec in caplog.records
|
| 230 |
+
if rec.levelno >= logging.WARNING
|
| 231 |
+
)
|
| 232 |
+
|
| 233 |
+
|
| 234 |
+
# ---------------------------------------------------------------------------
|
| 235 |
+
# Bug 3 — Cohérence runner/rapport : empty hypothesis → CER 1.0 dans DocumentResult
|
| 236 |
+
# ---------------------------------------------------------------------------
|
| 237 |
+
|
| 238 |
+
class TestRunnerDocumentResultCohérence:
|
| 239 |
+
"""Le DocumentResult doit stocker CER=1.0 pour une hypothèse vide."""
|
| 240 |
+
|
| 241 |
+
def test_empty_hypothesis_stored_as_cer_one(self):
|
| 242 |
+
"""_compute_document_result avec text="" → metrics.cer = 1.0."""
|
| 243 |
+
from picarones.core.runner import _compute_document_result
|
| 244 |
+
from picarones.engines.base import EngineResult
|
| 245 |
+
|
| 246 |
+
ocr_result = EngineResult(
|
| 247 |
+
engine_name="TestEngine",
|
| 248 |
+
image_path="fake.png",
|
| 249 |
+
text="", # ← sortie vide
|
| 250 |
+
duration_seconds=1.0,
|
| 251 |
+
error=None, # ← pas d'erreur technique
|
| 252 |
+
)
|
| 253 |
+
|
| 254 |
+
doc_result = _compute_document_result(
|
| 255 |
+
doc_id="doc1",
|
| 256 |
+
image_path="fake.png",
|
| 257 |
+
ground_truth="Bonjour le monde",
|
| 258 |
+
ocr_result=ocr_result,
|
| 259 |
+
char_exclude=None,
|
| 260 |
+
)
|
| 261 |
+
|
| 262 |
+
assert doc_result.metrics.cer == pytest.approx(1.0), (
|
| 263 |
+
f"CER attendu 1.0 pour hypothèse vide, obtenu {doc_result.metrics.cer}"
|
| 264 |
+
)
|
| 265 |
+
assert doc_result.metrics.error is None, (
|
| 266 |
+
"L'erreur ne devrait pas être renseignée — c'est une hypothèse vide, pas une erreur technique"
|
| 267 |
+
)
|
| 268 |
+
|
| 269 |
+
def test_engine_error_also_gives_cer_one(self):
|
| 270 |
+
"""EngineResult avec error → metrics.cer = 1.0 (comportement existant)."""
|
| 271 |
+
from picarones.core.runner import _compute_document_result
|
| 272 |
+
from picarones.engines.base import EngineResult
|
| 273 |
+
|
| 274 |
+
ocr_result = EngineResult(
|
| 275 |
+
engine_name="TestEngine",
|
| 276 |
+
image_path="fake.png",
|
| 277 |
+
text="",
|
| 278 |
+
duration_seconds=0.0,
|
| 279 |
+
error="Moteur en erreur",
|
| 280 |
+
)
|
| 281 |
+
|
| 282 |
+
doc_result = _compute_document_result(
|
| 283 |
+
doc_id="doc1",
|
| 284 |
+
image_path="fake.png",
|
| 285 |
+
ground_truth="Bonjour le monde",
|
| 286 |
+
ocr_result=ocr_result,
|
| 287 |
+
char_exclude=None,
|
| 288 |
+
)
|
| 289 |
+
|
| 290 |
+
assert doc_result.metrics.cer == pytest.approx(1.0)
|
| 291 |
+
assert doc_result.metrics.error is not None
|