DSPy : le banger qui booste tes prompts et ton RAG

pip install dspy

import dspy
from dspy import Example, LM, Module, Prediction, ChainOfThought, Signature, InputField, OutputField
from dspy.teleprompt import GEPA

# Ici tu déclare ton LLM et les paramètres d'appel
lm = LM(
    model="openai/gpt-oss",
    api_base="https://api-openai-compatible",
    api_key="sk-S3CR3T",
    api_version="v1",
    temperature=0.7,
    max_tokens=1800,
    cache=False, # le cache ici peut être utile pour accélérer les tests
)

dspy.settings.configure(lm=lm)

# Ensuite on va définir notre signature, avec une phrase en entrée et un sentiment en sortie.
# La docstring peut aider à guider le modèle sur la tâche.
class SentimentSignature(Signature):
    """Classify the sentiment of the sentence into one word: positive, negative, or neutral."""

    phrase: str = InputField()
    sentiment: str = OutputField(desc="positive | negative | neutral")


# On fait notre module avec une chaine de pensée basé sur la signature.
class SimpleSentiment(Module):
    def __init__(self):
        super().__init__()
        self.classify = ChainOfThought(SentimentSignature)

    def forward(self, phrase):
        return self.classify(phrase=phrase)


# On définit les jeux de données d'entraînement et de validation.
# Ici on a juste quelques exemples pour montrer le principe, mais tu peux en mettre beaucoup plus pour de meilleurs résultats.
trainset = [
    Example(phrase="J'adore ce film !",                 sentiment="positive").with_inputs("phrase"),
    Example(phrase="Vraiment nul, à fuir.",             sentiment="negative").with_inputs("phrase"),
    Example(phrase="C'est correct.",                    sentiment="neutral").with_inputs("phrase"),
    Example(phrase="Incroyable, je recommande !",      sentiment="positive").with_inputs("phrase"),
    Example(phrase="Déçu, je m'attendais à mieux.",    sentiment="negative").with_inputs("phrase"),
    Example(phrase="Pas mal du tout.",                  sentiment="positive").with_inputs("phrase"),
    Example(phrase="Rien de spécial.",                  sentiment="neutral").with_inputs("phrase"),
    Example(phrase="Magnifique paysage !",              sentiment="positive").with_inputs("phrase"),
]

# Un petit jeu de validation pour que GEPA puisse évaluer les différentes versions du prompt.
valset = [
    Example(phrase="Service horrible.",                 sentiment="negative").with_inputs("phrase"),
    Example(phrase="Bof...",                            sentiment="neutral").with_inputs("phrase"),
    Example(phrase="Un pur chef-d'œuvre !",             sentiment="positive").with_inputs("phrase"),
    Example(phrase="Franchement décevant.",             sentiment="negative").with_inputs("phrase"),
]


# Ici on définit la métrique d'évaluation qui va permettre à GEPA de juger les différentes versions du prompt.
# On compare simplement le sentiment prédit avec le sentiment attendu, et on donne un feedback détaillé en cas d'erreur
# pour aider GEPA à comprendre ce qui ne va pas.
def sentiment_metric(
    gold: Example,
    pred: Prediction,
    trace=None,
    pred_name=None,
    pred_trace=None,
) -> Prediction:
    gold_label = gold.sentiment.lower().strip()
    pred_label = getattr(pred, "sentiment", "").lower().strip() if pred else ""

    is_correct = gold_label == pred_label

    if is_correct:
        feedback = "Correct ✓"
        score = 1.0
    else:
        feedback = (
            f"Error: expected '{gold_label}', got '{pred_label}'.\n"
            f"→ Model likely missed tone, irony, negation, or strong sentiment words."
        )
        score = 0.0

    return Prediction(score=score, feedback=feedback)


# On crée notre optimiseur GEPA en lui fournissant la metric d'évaluation, le LM, et quelques paramètres pour guider l'optimisation.
# Ici on est parti sur du light pour tester rapidement.
optimizer = GEPA(
    metric=sentiment_metric,
    auto="light",
    reflection_lm=lm,
    reflection_minibatch_size=3,
    num_threads=6,
    track_stats=True,
    add_format_failure_as_feedback=True,
    # seed=42, # tu peux fixer une seed pour la reproductibilité, mais c'est pas obligatoire
)


# On lance l'optimisation de GEPA sur notre module et nos datasets d'entraînement et de validation.
print("Starting GEPA optimization (may take 2–15 minutes depending on gateway load)...")
optimized = optimizer.compile(
    SimpleSentiment(),
    trainset=trainset,
    valset=valset,
)
print("\nOptimization finished.\n")


# On évalue les performances du modèle optimisé sur un ensemble de phrases de test.
# Ici on utilise quelques phrases de test pour voir comment le modèle généralise.
test_phrases = [
    "C'est génial, je suis fan !",
    "Franchement bof, décevant.",
    "Ça va, rien à dire.",
    "Le pire film de l'année.",
    "Absolument sublime, un chef-d'œuvre.",
    "Niveau zéro, honteux.",
    "Plutôt sympa en fait.",
    "Je ne m'attendais pas à autant aimer.",
]

# On affiche les résultats des prédictions pour chaque phrase de test.
print("Test results:\n")
for phrase in test_phrases:
    prediction = optimized(phrase=phrase)
    sentiment = getattr(prediction, "sentiment", "—").strip()
    print(f"{phrase:<50} → {sentiment}")


# Enfin, on peut aussi jeter un œil au prompt optimisé que GEPA a trouvé pour notre tâche.
print("\nOptimized instruction found by GEPA:")
for name, pred in optimized.named_predictors():
    print("================================")
    print(f"Predictor: {name}")
    print("================================")
    print(pred.signature.instructions)

Task: Sentiment classification of a single short sentence or phrase.

**Input format**
- The user will provide a block labeled `### phrase` containing the sentence to analyze.
- The sentence may be in any language (e.g., French, English, etc.). You may translate it internally if that helps you judge the sentiment, but you must keep the original wording in your explanation.

**Output format**
Your response must contain two markdown sections, exactly as shown below:

```
### reasoning
<brief explanation (1‑2 sentences) of why the sentence is positive, negative, or neutral>

### sentiment
<one word: positive, negative, or neutral>
```

- The word in the `### sentiment` line must be **lower‑case** and one of the three allowed labels.
- No additional text, emojis, or formatting may appear after the sentiment word.

**How to decide the label**

1. **Positive** – The sentence expresses a clear favorable attitude, pleasure, admiration, approval, excitement, or any other overtly upbeat emotion. Typical cues include:
- Positive adjectives or adverbs (e.g., *magnifique, great, wonderful, happy*).
- Exclamation marks that reinforce enthusiasm.
- Explicit statements of liking, love, gratitude, or endorsement.

2. **Negative** – The sentence conveys dissatisfaction, dislike, anger, sadness, criticism, disappointment, or any other unfavorable feeling. Look for:
- Negative adjectives or adverbs (e.g., *terrible, awful, horrible, sad*).
- Words of complaint, rejection, or condemnation.
- Expressions of frustration, hurt, or hostility.

3. **Neutral** – The sentence is factual, descriptive, or merely acknowledges something without showing strong affect. It includes:
- Simple statements of fact or correctness (e.g., *C’est correct.*, *Il fait 20 °C.*).
- Mild assessments that are neither clearly approving nor disapproving.
- Sentences where any sentiment is ambiguous, mixed, or overridden by sarcasm/irony that neutralizes the literal meaning.

**Special considerations**

- **Negation**: Phrases like “not good”, “no thanks”, “nothing special” flip the sentiment of the word they negate. Treat the overall sentiment as negative if the negation creates an unfavorable meaning, or neutral if it merely softens a positive statement.
- **Sarcasm / Irony**: Detect when a positive‑looking word is used sarcastically (e.g., “Great, another meeting”). In such cases, label the sentiment as **negative**. If sarcasm merely dampens enthusiasm without a clear negative tone, choose **neutral**.
- **Mild positivity/negativity**: If the emotional intensity is very weak (e.g., “It’s okay”, “Not bad”), treat it as **neutral** unless the context clearly leans toward a positive or negative judgment.
- **Exclamation marks**: They amplify sentiment but do not alone determine it. Combine them with the lexical content.
- **Language‑specific cues**: Be aware of idiomatic expressions in the source language that carry sentiment (e.g., French *super*, *nul*, *pas mal*). Translate mentally if needed, but keep the original wording in the reasoning.

**General strategy**
1. Read the sentence carefully.
2. Identify sentiment‑bearing words, punctuation, and any negation or sarcasm.
3. Decide whether the overall affect is clearly positive, clearly negative, or otherwise neutral.
4. Write a concise reasoning sentence explaining the key cue(s) you used.
5. Output the sentiment label in the required format.

Follow these rules strictly to ensure consistent, reproducible classifications.

class RAGSignature(Signature):
    """Answer the question based on the provided context."""

    question: str = InputField()
    context: str = InputField()
    answer: str = OutputField(desc="quality score or feedback")