Bing Chat: GPT-4 Powered Search

I wrote “Bing Chat: GPT-4 Powered Search” to share practical, production-minded guidance on this topic.

Architecture: Search + LLM

class BingChatArchitecture:
    """Conceptual architecture of Bing Chat."""

    async def process_query(self, user_query: str) -> dict:
        """Process user query with search augmentation."""

        # Step 1: Determine if search is needed
        needs_search = await self._classify_query(user_query)

        if needs_search:
            # Step 2: Generate search queries
            search_queries = await self._generate_search_queries(user_query)

            # Step 3: Execute searches
            search_results = await self._search(search_queries)

            # Step 4: Generate response with search context
            response = await self._generate_with_context(
                user_query,
                search_results
            )
        else:
            # Direct response without search
            response = await self._generate_direct(user_query)

        return response

    async def _classify_query(self, query: str) -> bool:
        """Determine if query needs web search."""
        # Queries about current events, specific facts, recent data
        current_indicators = [
            "today", "latest", "current", "now", "2023",
            "price", "weather", "news", "stock"
        ]
        return any(ind in query.lower() for ind in current_indicators)

    async def _generate_search_queries(self, user_query: str) -> list[str]:
        """Generate optimized search queries."""
        prompt = f"""Generate 3 search queries to answer this question.

Question: {user_query}

Return queries that:
1. Cover different aspects
2. Use effective search terms
3. Target authoritative sources

Return as JSON array."""

        response = await self.client.chat_completion(
            model="gpt-4",
            messages=[{"role": "user", "content": prompt}],
            temperature=0.3
        )

        import json
        try:
            return json.loads(response.content)
        except:
            return [user_query]

    async def _generate_with_context(
        self,
        query: str,
        search_results: list[dict]
    ) -> dict:
        """Generate response with search context."""

        # Format search results
        context = self._format_search_results(search_results)

        prompt = f"""Answer this question using the search results.

Question: {query}

Search Results:
{context}

Instructions:
- Synthesize information from multiple sources
- Cite sources using [1], [2], etc.
- Indicate when information might be outdated
- Be direct and helpful"""

        response = await self.client.chat_completion(
            model="gpt-4",
            messages=[{"role": "user", "content": prompt}]
        )

        return {
            "answer": response.content,
            "sources": [r["url"] for r in search_results[:5]],
            "search_used": True
        }

Conversation Modes

Bing Chat offers three modes:

from enum import Enum

class BingChatMode(Enum):
    CREATIVE = "creative"   # More imaginative, longer responses
    BALANCED = "balanced"   # Balance of accuracy and creativity
    PRECISE = "precise"     # Factual, concise, search-focused

class BingChatModeSettings:
    """Settings for each conversation mode."""

    SETTINGS = {
        BingChatMode.CREATIVE: {
            "temperature": 0.9,
            "max_tokens": 2000,
            "search_emphasis": 0.3,
            "style": "engaging and creative"
        },
        BingChatMode.BALANCED: {
            "temperature": 0.5,
            "max_tokens": 1000,
            "search_emphasis": 0.5,
            "style": "helpful and balanced"
        },
        BingChatMode.PRECISE: {
            "temperature": 0.2,
            "max_tokens": 500,
            "search_emphasis": 0.8,
            "style": "factual and concise"
        }
    }

    @classmethod
    def get_system_prompt(cls, mode: BingChatMode) -> str:
        settings = cls.SETTINGS[mode]
        return f"""You are Bing, an AI assistant.
Mode: {mode.value}
Style: {settings['style']}

Guidelines:
- Cite sources for factual claims
- Be helpful and informative
- Acknowledge limitations
- Avoid harmful content"""

Building Similar Experiences

class SearchAugmentedLLM:
    """Build search-augmented LLM applications."""

    def __init__(self, client, search_api):
        self.client = client
        self.search = search_api

    async def answer_with_search(
        self,
        question: str,
        max_sources: int = 5
    ) -> dict:
        """Answer question using web search."""

        # Search
        results = await self.search.search(question)
        top_results = results[:max_sources]

        # Fetch and extract content
        sources_content = []
        for result in top_results:
            content = await self._extract_content(result["url"])
            sources_content.append({
                "title": result["title"],
                "url": result["url"],
                "content": content[:2000]
            })

        # Generate answer
        answer = await self._generate_answer(question, sources_content)

        return {
            "answer": answer,
            "sources": [{"title": s["title"], "url": s["url"]} for s in sources_content]
        }

    async def _extract_content(self, url: str) -> str:
        """Extract main content from URL."""
        # Implementation depends on your needs
        # Could use trafilatura, newspaper3k, etc.
        pass

    async def _generate_answer(
        self,
        question: str,
        sources: list[dict]
    ) -> str:
        """Generate answer from sources."""

        sources_text = "\n\n".join([
            f"[{i+1}] {s['title']}\n{s['content']}"
            for i, s in enumerate(sources)
        ])

        prompt = f"""Answer this question using the provided sources.

Question: {question}

Sources:
{sources_text}

Answer citing sources as [1], [2], etc."""

        response = await self.client.chat_completion(
            model="gpt-4",
            messages=[{"role": "user", "content": prompt}]
        )

        return response.content

Key Differences from ChatGPT

Feature	ChatGPT	Bing Chat
Knowledge	Cutoff date	Live search
Sources	None	Cited links
Images	Text only	Can generate
Conversation	Unlimited	Limited turns
Modes	One	Creative/Balanced/Precise

Bing Chat represents the future of search: not just finding links, but synthesizing answers from the web with AI understanding.\n\n## Takeaways\n\nAdd a concise, personal takeaway and recommended next steps here.\n