Mitigation: Grounding, Self-Check, Citation

from anthropic import Anthropic

client = Anthropic()

def ask_with_cot(question: str) -> dict:
    """
    Structured chain-of-thought prompt that forces explicit reasoning
    before committing to an answer.
    """
    cot_prompt = f"""
Answer the following question using this exact structure:

QUESTION: {question}

STEP 1 - Identify what is being asked:
[Write one sentence about what the question requires]

STEP 2 - List what I know with confidence:
[Bullet each fact you are confident about, with a note if you are uncertain]

STEP 3 - Identify what I am uncertain about:
[Explicitly list any gaps in your knowledge]

STEP 4 - Reason through the answer:
[Work through the logic step by step]

STEP 5 - Final answer:
[State your answer clearly, qualifying anything uncertain]

CONFIDENCE: [high / medium / low] — [one sentence explaining why]
"""
    
    response = client.messages.create(
        model="claude-sonnet-4-6",
        max_tokens=800,
        messages=[{"role": "user", "content": cot_prompt}]
    )
    
    return {
        "question": question,
        "cot_response": response.content[0].text,
        "technique": "chain-of-thought"
    }

result = ask_with_cot("What is the recommended daily protein intake for adults?")
print(result["cot_response"])

import anthropic
from collections import Counter

client = anthropic.Anthropic()

def self_consistency_check(
    question: str,
    num_samples: int = 5,
    temperature: float = 0.7
) -> dict:
    """
    Generate multiple answers at non-zero temperature.
    Majority vote on the final answer increases reliability.
    """
    answers = []
    
    for _ in range(num_samples):
        response = client.messages.create(
            model="claude-sonnet-4-6",
            max_tokens=200,
            temperature=temperature,
            messages=[{
                "role": "user",
                "content": f"Answer in one sentence: {question}"
            }]
        )
        answers.append(response.content[0].text.strip())
    
    # Count frequency of each unique answer (simplified — real implementation
    # would use semantic clustering, not exact string matching)
    frequency = Counter(answers)
    most_common = frequency.most_common(1)[0]
    agreement_rate = most_common[1] / num_samples
    
    return {
        "question": question,
        "all_answers": answers,
        "consensus_answer": most_common[0],
        "agreement_rate": agreement_rate,
        "confidence": "high" if agreement_rate >= 0.8 else (
            "medium" if agreement_rate >= 0.6 else "low"
        ),
        "flag_for_review": agreement_rate < 0.6
    }

result = self_consistency_check("In what year did Python 3.0 release?", num_samples=5)
print(f"Consensus: {result['consensus_answer']}")
print(f"Agreement: {result['agreement_rate']:.0%}")
print(f"Flag for review: {result['flag_for_review']}")

UNCERTAINTY_SYSTEM_PROMPT = """
You are a helpful assistant with one strict rule:

When you are uncertain about a fact, you MUST use one of these markers:
- [UNCERTAIN]: I believe this is correct but am not fully confident
- [VERIFY]: Please verify this claim from a primary source
- [UNKNOWN]: I do not have reliable information about this
- [ESTIMATED]: This is an approximation, not a precise figure

Never state something you are uncertain about without the appropriate marker.
If a question is outside your knowledge, say so directly rather than guessing.
"""

def ask_with_uncertainty(question: str) -> str:
    response = client.messages.create(
        model="claude-sonnet-4-6",
        max_tokens=400,
        system=UNCERTAINTY_SYSTEM_PROMPT,
        messages=[{"role": "user", "content": question}]
    )
    return response.content[0].text

answer = ask_with_uncertainty("What is the population of Oslo, Norway?")
print(answer)
# "Oslo has a population of approximately 700,000 people [ESTIMATED — 
#  this figure is from my training data and may not reflect current numbers.
#  Please verify with Statistics Norway (ssb.no) for the current figure.]"

def rag_with_mandatory_quotes(question: str, retrieved_chunks: list[dict]) -> dict:
    """
    Forces the model to include verbatim quotes from source documents.
    Verbatim quotes are verifiable — paraphrases can drift.
    """
    context = "\n\n".join([
        f"[CHUNK {i+1} | Source: {c['source']} | Page: {c.get('page', 'N/A')}]\n{c['text']}"
        for i, c in enumerate(retrieved_chunks)
    ])
    
    system = """
You are a document Q&A assistant. Strict rules:
1. You MUST include at least one verbatim quote from the provided chunks.
   Format quotes as: > "exact text here" [Chunk N, Source: filename]
2. Only answer from the provided document chunks.
3. If chunks don't contain the answer, say: "This is not covered in the provided documents."
4. Do not paraphrase if you can quote directly.
"""
    
    response = client.messages.create(
        model="claude-sonnet-4-6",
        max_tokens=500,
        system=system,
        messages=[{
            "role": "user",
            "content": f"Document chunks:\n{context}\n\nQuestion: {question}"
        }]
    )
    
    return {
        "answer": response.content[0].text,
        "sources_provided": [c["source"] for c in retrieved_chunks]
    }

def rag_with_excerpts(question: str, store) -> dict:
    """
    Returns both the answer AND the raw retrieved excerpts.
    Users or downstream systems can verify the answer against the excerpt.
    """
    chunks = store.search(question, top_k=2)
    answer_result = rag_with_mandatory_quotes(question, chunks)
    
    return {
        "answer": answer_result["answer"],
        "supporting_excerpts": [
            {
                "source": c["source"],
                "excerpt": c["text"][:300] + "..." if len(c["text"]) > 300 else c["text"],
                "relevance_score": round(c["score"], 3)
            }
            for c in chunks
        ]
    }

from transformers import pipeline
import re

# Load a cross-encoder NLI model — better than bi-encoder for this task
nli_pipeline = pipeline(
    "text-classification",
    model="cross-encoder/nli-deberta-v3-small",
    device=-1  # CPU; use 0 for GPU
)

def split_into_sentences(text: str) -> list[str]:
    """Basic sentence splitter."""
    sentences = re.split(r'(?<=[.!?])\s+', text)
    return [s.strip() for s in sentences if len(s.strip()) > 10]

def nli_faithfulness_check(
    model_answer: str,
    retrieved_context: str,
    entailment_threshold: float = 0.5
) -> dict:
    """
    Check each sentence in the model's answer against the retrieved context.
    Flag sentences that are not entailed by the context.
    
    Returns a faithfulness report with per-sentence verdicts.
    """
    sentences = split_into_sentences(model_answer)
    results = []
    
    for sentence in sentences:
        # NLI: does context entail this sentence?
        nli_input = f"{retrieved_context} [SEP] {sentence}"
        prediction = nli_pipeline(nli_input, truncation=True, max_length=512)
        
        label = prediction[0]["label"].upper()
        score = prediction[0]["score"]
        
        verdict = "FAITHFUL"
        if label == "CONTRADICTION":
            verdict = "CONTRADICTS_CONTEXT"
        elif label == "NEUTRAL" and score > entailment_threshold:
            verdict = "NOT_IN_CONTEXT"
        
        results.append({
            "sentence": sentence,
            "nli_label": label,
            "nli_score": round(score, 3),
            "verdict": verdict
        })
    
    flagged = [r for r in results if r["verdict"] != "FAITHFUL"]
    overall_faithful = len(flagged) == 0
    
    return {
        "overall_faithful": overall_faithful,
        "faithfulness_score": 1.0 - (len(flagged) / max(len(sentences), 1)),
        "flagged_count": len(flagged),
        "sentence_results": results,
        "action": "PASS" if overall_faithful else "REVIEW_FLAGGED_SENTENCES"
    }


def rag_with_nli_guard(
    question: str,
    retrieved_chunks: list[dict],
    faithfulness_threshold: float = 0.8
) -> dict:
    """
    Full pipeline: retrieve → generate → NLI check → return or flag.
    """
    context = " ".join([c["text"] for c in retrieved_chunks])
    
    # Generate answer
    response = client.messages.create(
        model="claude-sonnet-4-6",
        max_tokens=400,
        system="Answer only from the provided context.",
        messages=[{
            "role": "user",
            "content": f"Context:\n{context}\n\nQuestion: {question}"
        }]
    )
    answer = response.content[0].text
    
    # NLI check
    faithfulness = nli_faithfulness_check(answer, context)
    
    if faithfulness["faithfulness_score"] < faithfulness_threshold:
        return {
            "answer": answer,
            "safe_to_show": False,
            "faithfulness": faithfulness,
            "recommendation": "Flag for human review before showing to user"
        }
    
    return {
        "answer": answer,
        "safe_to_show": True,
        "faithfulness": faithfulness
    }

import json

def generate_with_confidence(question: str, context: str) -> dict:
    """
    Ask the model to self-report confidence AND provide supporting evidence.
    Then validate that the claimed confidence matches the evidence.
    """
    system = """
You are a careful analyst. For every answer, respond with this JSON structure:
{
  "answer": "your answer",
  "confidence": "high|medium|low",
  "confidence_reasons": {
    "supporting": ["reason 1", "reason 2"],
    "against": ["uncertainty 1", "uncertainty 2"]
  },
  "would_recommend_verification": true/false,
  "verification_source": "where to verify this (URL, document name, etc.)"
}
Base confidence on: source quality, recency, specificity, and your certainty.
"""
    
    response = client.messages.create(
        model="claude-sonnet-4-6",
        max_tokens=500,
        system=system,
        messages=[{
            "role": "user",
            "content": f"Context:\n{context}\n\nQuestion: {question}"
        }]
    )
    
    raw = response.content[0].text.strip()
    if raw.startswith("```"):
        lines = raw.split("\n")
        raw = "\n".join(lines[1:-1])
    
    try:
        result = json.loads(raw)
        # Map confidence to numeric score for downstream use
        confidence_map = {"high": 0.9, "medium": 0.6, "low": 0.3}
        result["confidence_score"] = confidence_map.get(result.get("confidence", "low"), 0.3)
        result["show_to_user"] = result["confidence_score"] >= 0.6
        return result
    except json.JSONDecodeError:
        return {
            "answer": raw,
            "confidence": "unknown",
            "confidence_score": 0.0,
            "show_to_user": False,
            "parse_error": True
        }

# Usage
result = generate_with_confidence(
    question="What is the half-life of aspirin in the human body?",
    context="Aspirin (acetylsalicylic acid) has a short half-life of approximately 15-20 minutes..."
)
print(f"Confidence: {result.get('confidence')} ({result.get('confidence_score')})")
print(f"Show to user: {result.get('show_to_user')}")

class HallucinationMitigationPipeline:
    """
    Combines all four mitigation layers into a single pipeline.
    """
    
    def __init__(self, vector_store, use_nli: bool = True):
        self.store = vector_store
        self.use_nli = use_nli
        self.client = Anthropic()
    
    def run(self, question: str) -> dict:
        # Layer 1: Retrieve with relevance check
        chunks = self.store.search(question, top_k=3)
        MIN_RELEVANCE = 0.3
        relevant = [c for c in chunks if c["score"] >= MIN_RELEVANCE]
        
        if not relevant:
            return {
                "answer": "Insufficient information in knowledge base.",
                "confidence_score": 0.0,
                "passed_all_checks": False,
                "failure_reason": "no_relevant_context"
            }
        
        context = " ".join([c["text"] for c in relevant])
        
        # Layer 2: Generate with CoT and uncertainty markers
        cot_response = ask_with_cot(question)
        
        # Layer 3: NLI faithfulness check
        if self.use_nli:
            faithfulness = nli_faithfulness_check(
                cot_response["cot_response"], context
            )
            if not faithfulness["overall_faithful"]:
                return {
                    "answer": cot_response["cot_response"],
                    "confidence_score": 0.2,
                    "passed_all_checks": False,
                    "failure_reason": "nli_faithfulness_check_failed",
                    "faithfulness_report": faithfulness
                }
        
        # Layer 4: Confidence scoring
        confidence = generate_with_confidence(question, context)
        
        return {
            "answer": confidence.get("answer"),
            "confidence_score": confidence.get("confidence_score", 0.0),
            "confidence_reasons": confidence.get("confidence_reasons", {}),
            "passed_all_checks": True,
            "sources": [c["metadata"]["source"] for c in relevant],
            "show_to_user": confidence.get("show_to_user", False)
        }