mkmanish/truthmark-api

1# predictor.py2 3import numpy as np4from model_loader import infer, tokenizer, MAX_LENGTH5 6 7def compute_breakdown_scores(text):8    tokens = [t.lower() for t in text.split() if t.isalpha() or t.isalnum()]9    n = max(1, len(tokens))10 11    trigrams = [' '.join(tokens[i:i+3]) for i in range(max(0, n-2))]12    repetition = 1 - (len(set(trigrams)) / max(1, len(trigrams)))13 14    types = len(set(tokens))15    ttr = types / n16    vocab_richness = 1 - ttr17 18    sentences = [s.strip() for s in text.replace('?', '.').replace('!', '.').split('.') if s.strip()]19    if len(sentences) <= 1:20        synt_var = 0.021    else:22        lens = [len(s.split()) for s in sentences]23        synt_var = 1 - (np.std(lens) / (np.mean(lens) + 1e-6))24 25    return {26        "repetition": round(np.clip(repetition, 0, 1) * 100, 2),27        "vocab_richness": round(np.clip(vocab_richness, 0, 1) * 100, 2),28        "syntactic_variation": round(np.clip(synt_var, 0, 1) * 100, 2)29    }30 31def verdict_from_prob(prob):32    score = round(prob * 100, 2)33    if score >= 70:34        return score, "Likely AI-Generated", "High", "red"35    elif score >= 40:36        return score, "Mixed/Uncertain", "Medium", "yellow"37    else:38        return score, "Likely Human-Created", "Low", "green"39 40def analyze_text(text: str):41    enc = tokenizer(42        [text],43        padding="max_length",44        truncation=True,45        max_length=128,46        return_tensors="tf"47    )48 49    outputs = infer(50        input_ids=enc["input_ids"],51        attention_mask=enc["attention_mask"]52    )53 54    raw_output = float(outputs[list(outputs.keys())[0]][0][0])55 56 57    # Convert safely to float and clamp58    prob_ai = float(np.clip(raw_output, 0.0, 1.0))59 60    score, verdict, confidence, badge_color = verdict_from_prob(prob_ai)61    breakdown = compute_breakdown_scores(text)62 63    return {64        "overall_ai_score": score,65        "overall_human_score": round(100 - score, 2),66        "verdict": verdict,67        "confidence": confidence,68        "badge_color": badge_color,69        70        "breakdown": {71            "Perplexity_proxy_Repetition": breakdown["repetition"],72            "Vocabulary_Richness_proxy": breakdown["vocab_richness"],73            "Syntactic_Variation_proxy": breakdown["syntactic_variation"]74        },75        "model_note": "This result is based on a probabilistic AI-vs-human classifier trained on multiple datasets.",76 77 78        "limitations": [79            "This system cannot guarantee 100% accuracy and should not be used as the only source to judge content authenticity ",80            "Text written by humans with very formal or repetitive style may sometimes appear AI-generated.",81            "AI-generated text that is heavily edited by humans may not be detected correctly.",82            "This tool does not identify which AI tool was used, only the likelihood of AI involvement."83        ]84    }85