Become the Highest Paid Engineer at Your Company/

...

/

Retrieval Beats Hallucination

Retrieval Beats Hallucination

Learn why LLMs hallucinate, and how retrieval plus llm.txt files ground answers in reality.

We'll cover the following...

At the Staff+ level, you’re responsible for building AI systems that are reliable, scalable, and safe to depend on.

One of the most common—and most dangerous—failure modes in AI systems? Hallucination.

Even the most state-of-the-art models hallucinate. It’s simply baked into how they are trained. During training, they learn that saying “I don’t know” rarely earns a reward, so instead, they get very good at guessing in a confident way.

Hallucinations usually fall into two categories:

  • Extrinsic hallucinations: Fabrications about the outside world (e.g., a refund policy that never existed).

  • Intrinsic hallucinations: Slips on mechanical tasks where the model produces a fluent but incorrect answer (e.g., counting the “b”s in “blueberry”).

The fix for both is the same: retrieval-augmented generation (RAG)—and it’s your best line of defense.

The power of RAG

RAG lets a model look up facts from a curated knowledge base when a question is asked,Or in ML terms: inference time. instead of relying only on what it memorizes. It retrieves the most relevant documents and then generates an answer grounded in those sources—often with quotes or citations.

Instead of asking the model to “remember,” you supply authoritative knowledge at runtime. That flips it from storyteller to fact-checker.

Let’s say you ask GPT, “What’s MCP?”

  • Without retrieval, the model might answer: “MCP was introduced by Google in 2019 as part of TensorFlow Extended.” 

    • This is polished, but completely wrong. MCP isn’t tied to Google or TensorFlow, and there was never a “2019 release.”

  • With retrieval, the model answers: “According to the MCP docs, MCP is an open-source standard that lets AI applications like Claude or ChatGPT connect to external systems such as databases, files, and tools. The docs describe it as a ‘USB-C for AI,’ but they don’t mention a release date or a single company owner.”

    • Much better.

Same model, different behavior—the only differentiating factor is retrieval.

A simple retrieval workflow

’Staying in tune with the MCP example, applying RAG would mean not chatting with the model when exploring a new technology.

Instead, you’d create a lightweight briefing file:

  1. Generate a file (llm.txt) from authoritative sources (docs, specs, APIs). Think of it as the model’s briefing packet.

  2. Feed it into your model (or index it with your retrieval layer).

  3. In prompts, explicitly say: “Answer only using the provided CONTEXT.”

Tools like Firecrawl’s create-llmstxt-py automate the scraping and summarizing.  

Setup requires two API keys:

  • Firecrawl API key

  • OpenAI API key

And setup is simple:

git clone https://github.com/firecrawl/create-llmstxt-py
cd create-llmstxt-py
pip install -r requirements.txt
cp .env.example .env  # then add your API keys
python generate-llmstxt.py https://example.com --max-urls 50

This outputs llm.txt (briefing) and llm-full.txt (detailed version), which you can feed into your LLM of choice.

There are plenty of other RAG tools out there. A quick search will show many commercial options, but they’re often paid and pricy. Firecrawl's create-llmstxt-py is a great place to start.

Building a RAG pipeline

If you want to push beyond John’s LLM.txt workflow, the next level is building full RAG pipelines yourself. 

A RAG pipeline is the production-grade version of the simple workflow we did:

  • Docs are ingested, chunked, indexed, retrieved, re-ranked, cited, evaluated, and refreshed on schedule. 

  • It’s engineered with guardrails (abstains, schema outputs, eval metrics, observability).

With a local RAG setup, you can feed your models authoritative knowledge, deploy systems into production, and stop treating AI as a novelty tool.

Get a guided start on building RAG pipelines with our hands-on courses:

👉 Fundamentals of Retrieval-Augmented Generation with LangChain: Learn to apply LangChain to implement RAG pipeline and build a frontend app for your pipeline with Streamlit.

👉 Advanced RAG Techniques: Choosing the Right Approach: Go deeper with different RAG approaches, post-retrieval optimization methods, and designing RAG-based chatbots.



Ask