LCEL SearchAgent

A small Express + TypeScript service that answers questions using an LLM.
It automatically chooses between two paths:

• Direct mode: respond from the model only (no browsing)
• Web mode: search the web, open top pages, summarize them, then answer with source URLs

The goal is to keep answers short and beginner-friendly while returning structured JSON output.

What this API does

High Level Design--> When you send a query to the API:

1. The request is validated with Zod.
2. A router step decides the mode:
   • Uses direct for normal / stable questions
   • Switches to web if the query looks time-sensitive, numeric, price-related, trending, “latest”, etc.
3. Direct mode
   • Calls the chat model once
   • Returns { answer, sources: [] }
4. Web mode
   • Uses Tavily Search to get results
   • Opens the top pages (HTTP fetch)
   • Converts HTML to text and caps content length
   • Summarizes each page with the model
   • Composes a final answer using only those summaries
   • Returns { answer, sources: [url, ...] }
5. Final output is validated again with Zod, and repaired once if needed.

Tech stack

• Node.js + TypeScript
• Express
• LangChain (RunnableSequence, RunnableBranch)
• Zod (input/output validation)
• Tavily Search API (web search)
• html-to-text (cleaning page HTML)
• CORS configured for a single allowed frontend origin

API

POST /api/search

Request body

{
  "q": "your question here"
}

Success Response

{
  "answer": "string",
  "sources": ["https://example.com", "https://example2.com"]
}