AI Models for RFP Scraping & Summarization

AI Models for RFP Scraping & Summarization

To parse and summarize lengthy English RFPs into a reference site, we compared leading OpenRouter models (e.g. OpenAI GPT-4, Claude, Google Gemini) against Chinese-origin models (DeepSeek, Zhipu’s GLM-4.5, Baidu ERNIE, Alibaba Qwen). We evaluated token cost, context window, summarization/Q&A strength, search/reasoning features, function-calling/tool use, and throughput suitability.

Token Cost

Western LLMs tend to be expensive. For example, OpenAI’s GPT-4o charges USD $5.00 per 1M input tokens and $20.00 per 1M output tokens. In contrast, Chinese models are vastly cheaper or free: DeepSeek-V2 was priced at only ¥1 (~$0.14) per million tokens, and Alibaba and Baidu promote their models as 20–40× lower cost than Western alternatives. Google’s Gemini 2.5 (via Google Cloud) is moderately priced: Gemini 2.5 Flash (1M token context) costs about $0.30 per 1M input and $2.50 per 1M output. GPT-4’s smaller “mini” variant is cheaper (around $0.60/$2.40 per 1M) but with smaller context.

Context Length

For long RFPs, context size is critical. OpenAI’s base GPT-4 handles ~8K tokens (with a 32K variant available). Claude 3+ models support very large contexts (hundreds of thousands of tokens). Google Gemini 2.5 Flash offers a 1,048,576-token (1M) context window. Chinese models also support extended context: DeepSeek-R1 is open-source with 128K token context; Zhipu’s GLM-4.5 likewise supports 128K tokens; Baidu ERNIE 4.5 variants up to 128K context; Alibaba Qwen 2.5 models similarly allow 128K in / 8K out. These huge contexts can ingest entire RFP documents in one pass.

Summarization & Document Q&A

All large LLMs can perform summarization and Q&A, but quality varies. GPT-4/GPT-4o generally produce very fluent, accurate summaries. Gemini 2.5 also excels at coherent long-form summarization. Claude 3.7 Sonnet likewise yields high-quality summaries. Among Chinese models, reports show DeepSeek is “strong in English” and excels at structured tasks like coding, though it may be less creative than GPT-4. Zhipu’s GLM-4.5 is reported to achieve state-of-the-art reasoning and code generation (on par with GPT-4 on benchmarks) and is explicitly designed for complex tasks (agentic planning). Baidu ERNIE 4.5 (trained on 5.6T Chinese+English tokens) is tuned for “high fidelity in instruction-following, multi-turn conversation, long-form generation, and reasoning.” Evaluations show higher coherence/factuality on long text than previous versions. Alibaba’s Qwen 2.5 models are similarly improved on generating long text (able to produce 8K+ tokens of output). In practice, GPT-4 and Gemini likely give the most fluent summaries, but GLM-4.5 and ERNIE should handle English RFPs competently, and DeepSeek/Qwen offer solid performance at negligible cost.

Search and Reasoning

None of these LLMs inherently “search” the web, but some support retrieval augmentation. Gemini 2.5 Flash explicitly supports Grounding with Google Search, letting it query Google as a tool. GPT-4 can use integrated web search via tools/plugins (e.g. ChatGPT’s web browsing) or we can implement Retrieval-Augmented Generation (RAG) with a search API and vector DB. Claude 3.7 does not natively search, but Anthropic offers tools (e.g. Claude plug-in browsing). Zhipu’s GLM-4.5 is built for “agent applications” and supports code execution and tool use, so one can integrate custom search or database queries with it. Baidu’s ERNIE 4.5 documentation notes its suitability for search/RAG pipelines. In summary, Gemini and OpenAI provide the most direct search-integration support, but all can be used in a pipeline with retrieval (e.g. embedding RFP text, using vector DB or Google/Bing APIs, then summarizing results).

Function Calling / Tool Use

GPT-4 (and GPT-4o) have mature function-calling APIs (e.g. JSON-schema enforcement, plugins) for structured outputs. Gemini 2.5 Flash explicitly supports function calling (listed under “Capabilities”). Zhipu GLM-4.5, designed for autonomous agents, is reported to reliably use tools and APIs. Claude’s new 3.7 or 4+ lines are expected to support function calls. Chinese open models (DeepSeek/ERNIE/Qwen) do not natively include “tools”, but because they are open-source, one can build custom tool interfaces. In practice, GPT-4/Gemini lead in built-in function/tool support, while Chinese models would require more custom engineering.

High-Volume/Scalability

For heavy scraping and summarization, cost and rate limits matter. OpenAI’s models (especially GPT-4) are costly and have throughput limits. Gemini’s “Batch Mode” (50% off) can help large volumes. Claude (via Anthropic’s Claude Cloud) offers APIs with high throughput tiers. Chinese models (being free/open) excel here: DeepSeek and GLM-4.5 can be self-hosted or run on subsidized servers, giving unlimited queries at near-zero cost. Alibaba’s and Baidu’s cloud APIs for Qwen/ERNIE are comparatively cheap. Thus, for budget scraping, DeepSeek/GLM-4.5/ERNIE/Qwen stand out. For accuracy-heavy tasks, GPT-4/Gemini and GLM-4.5/ERNIE lead. For API flexibility, GPT-4 and Gemini (or Claude) are best, with robust tool support.

Summary Table

Recommendations by Use Case

• Budget scraping & generation: DeepSeek or GLM-4.5 (free) for long docs, or Gemini Flash-Lite for very low API cost.
• Best summaries: GPT-4/GPT-4o, then Gemini 2.5 Flash and Claude. Chinese: GLM-4.5 and ERNIE for solid English handling.
• API flexibility: GPT-4 and Gemini lead in function/tool use. Gemini offers built-in search integration. GLM-4.5 supports agent-like workflows.
• Scalability: Use Chinese open models (DeepSeek, ERNIE, Qwen) for bulk, reserve GPT-4/Gemini for final polish.

Integration Pipeline Suggestions

1. Retrieval: Index RFPs in a vector DB (e.g. Pinecone, Weaviate). Retrieve relevant passages.
2. LLM Stage: Summarize entire doc if context allows, otherwise chunked. Use cheaper models first, GPT-4/Gemini for final summaries.
3. Tool Use: For structured output (JSON), use GPT-4/Gemini function calling. Or chain models (one extracts, another formats).
4. Automation: Orchestrate with LangChain/LlamaIndex. Use headless browsers or APIs for scraping.
5. Search/QA: Combine LLM with search index to emulate “ChatGPT Search” or Perplexity.