LlamaIndex Web Scraping Integration — fastCRW [Firecrawl-Compatible]

Why LlamaIndex + fastCRW

LlamaIndex is the data framework for LLM applications — it connects your models to documents, APIs, and databases. fastCRW plugs in as the web scraping primitive. Instead of building your own scraper or using a heavyweight Firecrawl container, you write a lightweight LlamaIndex reader that calls fastCRW's REST API. The output is clean Markdown that feeds directly into chunking, embedding, and vector storage.

The pattern: ingest live web pages via fastCRW, let LlamaIndex chunk and embed them, query the vector store, and pass retrieved context to your LLM. Unlike batch scraping tools, fastCRW on-demand scraping keeps your RAG index fresh with zero infrastructure overhead.

Setup

1. Install LlamaIndex and dependencies.
2. Sign up at fastcrw.com for an API key.
3. Export FASTCRW_API_KEY in your shell.
4. Implement a custom LlamaIndex reader class that calls fastCRW.
5. Use the reader in your data loading and agent pipelines.

pip install llama-index requests
export FASTCRW_API_KEY="fcrw_..."

Code Example: Custom LlamaIndex Reader for fastCRW

Create a file fastcrw_reader.py:

import os
import requests
from typing import List, Any
from llama_index.core import Document
from llama_index.core.readers.base import BaseReader
from llama_index.core.schema import Document


class FastCRWReader(BaseReader):
    """
    A LlamaIndex reader that scrapes URLs via fastCRW REST API.
    Supports single-page scrape and recursive crawls.
    """

    def __init__(self, api_key: str | None = None, base_url: str = "https://fastcrw.com"):
        self.api_key = api_key or os.environ.get("FASTCRW_API_KEY")
        if not self.api_key:
            raise ValueError("FASTCRW_API_KEY not set")
        self.base_url = base_url

    def load_data(
        self,
        url: str,
        mode: str = "scrape",
        max_depth: int | None = None,
        formats: List[str] | None = None,
        **kwargs: Any,
    ) -> List[Document]:
        """
        Load and scrape URL(s) via fastCRW.

        Args:
            url: The target URL to scrape or crawl.
            mode: "scrape" (single page) or "crawl" (recursive).
            max_depth: For crawl mode, max crawl depth.
            formats: Output formats (default: ["markdown"]).
            **kwargs: Additional fastCRW parameters.

        Returns:
            List of LlamaIndex Document objects with scraped content.
        """
        if formats is None:
            formats = ["markdown"]

        headers = {
            "Authorization": f"Bearer {self.api_key}",
            "Content-Type": "application/json",
        }

        if mode == "scrape":
            # Single-page scrape
            payload = {
                "url": url,
                "formats": formats,
                **kwargs,
            }
            response = requests.post(
                f"{self.base_url}/api/v1/scrape",
                json=payload,
                headers=headers,
                timeout=60,
            )
            response.raise_for_status()
            data = response.json()

            doc_content = data["data"].get("markdown") or data["data"].get("html", "")
            return [
                Document(
                    text=doc_content,
                    metadata={
                        "url": url,
                        "status_code": data["data"].get("status_code"),
                        "load_time_ms": data["data"].get("load_time_ms"),
                    },
                )
            ]

        elif mode == "crawl":
            # Recursive crawl
            payload = {
                "url": url,
                "formats": formats,
                **kwargs,
            }
            if max_depth:
                payload["max_depth"] = max_depth

            response = requests.post(
                f"{self.base_url}/api/v1/crawl",
                json=payload,
                headers=headers,
                timeout=300,  # Crawls can take longer
            )
            response.raise_for_status()
            data = response.json()

            # Map crawl results to LlamaIndex documents
            documents = []
            for page in data["data"]:
                doc_content = page.get("markdown") or page.get("html", "")
                documents.append(
                    Document(
                        text=doc_content,
                        metadata={
                            "url": page.get("url"),
                            "status_code": page.get("status_code"),
                            "load_time_ms": page.get("load_time_ms"),
                        },
                    )
                )
            return documents

        else:
            raise ValueError(f"Unknown mode: {mode}. Use 'scrape' or 'crawl'.")

Usage Example: RAG Pipeline with fastCRW

from llama_index.core import VectorStoreIndex, Settings
from llama_index.embeddings.openai import OpenAIEmbedding
from llama_index.llms.openai import OpenAI
from fastcrw_reader import FastCRWReader

# Configure embedding and LLM
Settings.embed_model = OpenAIEmbedding(model="text-embedding-3-small")
Settings.llm = OpenAI(model="gpt-4o-mini")

# Instantiate the fastCRW reader
reader = FastCRWReader()

# Load documents from a website
documents = reader.load_data(
    url="https://docs.python.org/3/",
    mode="crawl",
    max_depth=2,  # Limit crawl depth for demo
)

# Create a vector index from scraped content
index = VectorStoreIndex.from_documents(documents)

# Query the index
query_engine = index.as_query_engine()
response = query_engine.query("What are Python's built-in functions?")
print(response)

Agent Tool Example: In-Loop Scraping

Expose fastCRW scraping as a LlamaIndex agent tool:

from llama_index.core.agent import ReActAgent
from llama_index.core.tools import FunctionTool
from llama_index.llms.openai import OpenAI
import requests
import os


def scrape_url(url: str) -> str:
    """
    Scrape a single URL via fastCRW and return Markdown content.
    """
    api_key = os.environ["FASTCRW_API_KEY"]
    response = requests.post(
        "https://fastcrw.com/api/v1/scrape",
        headers={"Authorization": f"Bearer {api_key}"},
        json={"url": url, "formats": ["markdown"]},
        timeout=60,
    )
    response.raise_for_status()
    data = response.json()
    return data["data"]["markdown"]


def search_web(query: str, limit: int = 5) -> str:
    """
    Search the web via fastCRW and return results.
    """
    api_key = os.environ["FASTCRW_API_KEY"]
    response = requests.post(
        "https://fastcrw.com/api/v1/search",
        headers={"Authorization": f"Bearer {api_key}"},
        json={"query": query, "limit": limit},
        timeout=30,
    )
    response.raise_for_status()
    results = response.json()["data"]["results"]
    return "\n".join(
        [f"- {r['title']}: {r['url']}" for r in results[:limit]]
    )


# Create tools
scrape_tool = FunctionTool.from_defaults(fn=scrape_url)
search_tool = FunctionTool.from_defaults(fn=search_web)

# Create agent
llm = OpenAI(model="gpt-4o-mini")
agent = ReActAgent.from_llm_and_tools(
    llm=llm,
    tools=[scrape_tool, search_tool],
    system_prompt="You are a research assistant. Use fastCRW tools to fetch live web content and answer questions based on current data.",
)

# Run agent reasoning loop
response = agent.chat("What are the latest Python 3.14 features?")
print(response)

Streaming Crawl Example

For large crawls, process pages as they arrive instead of batching:

import requests
import os
from llama_index.core import Document

def crawl_streaming(
    url: str,
    max_depth: int = 2,
    process_doc: callable = None,
) -> int:
    """
    Crawl a site via fastCRW and stream documents to a callback.

    Args:
        url: Starting URL.
        max_depth: Crawl depth limit.
        process_doc: Callback function (Document) -> None.

    Returns:
        Total pages crawled.
    """
    api_key = os.environ["FASTCRW_API_KEY"]
    payload = {
        "url": url,
        "formats": ["markdown"],
        "max_depth": max_depth,
    }

    response = requests.post(
        "https://fastcrw.com/api/v1/crawl",
        headers={"Authorization": f"Bearer {api_key}"},
        json=payload,
        timeout=300,
    )
    response.raise_for_status()
    data = response.json()

    count = 0
    for page in data["data"]:
        doc = Document(
            text=page.get("markdown", page.get("html", "")),
            metadata={"url": page.get("url")},
        )
        if process_doc:
            process_doc(doc)
        count += 1

    return count

When to Use This

• RAG systems with live web content — scrape documentation, blogs, and news into a vector index.
• Multi-document QA — crawl a website and build a searchable knowledge base.
• Data pipelines — ingest web content at intervals (daily, hourly) without manual crawler maintenance.
• Agent research loops — let agents scrape and summarize in real time during reasoning.
• Competitive intelligence — scrape multiple competitor sites and aggregate with LlamaIndex tools.
• Content curation — scrape RSS feeds and webpages, chunk, embed, and surface via semantic search.

Limits + Gotchas

• Crawl timeouts — large crawls can take 30+ seconds. Set appropriate request timeouts in your reader.
• Rate limiting — fastCRW enforces API rate limits. Implement exponential backoff if you hit 429 responses.
• Memory usage — storing all pages from a large crawl in memory can be expensive. Use streaming or pagination.
• Metadata mapping — fastCRW and LlamaIndex use different metadata schemas. Map custom fields explicitly in the Document constructor.
• Duplicate URLs — the crawl endpoint may return duplicates. Deduplicate by URL before indexing.
• JS rendering — static HTTP scraping is the default. For JS-heavy sites, request LightPanda or Chrome rendering (incurs cost).
• Authentication — fastCRW cannot authenticate automatically. Scrape public content or use session persistence for authenticated flows.

Performance Notes

• Single-page scrape: 833 ms median on HTTP content.
• Crawl speed: Depends on site complexity and depth. Budget 1–5 seconds per page.
• Embedding latency: Chunking and embedding 100 pages adds ~10–30 seconds with OpenAI embeddings.
• Vector store: LlamaIndex supports in-memory (demo), Chroma, Pinecone, and others. Choose based on scale.

Related

• LlamaIndex documentation
• fastCRW REST API docs
• LangChain integration
• Pydantic AI integration