Dear AI "Experts"

Please stop saying…

“LLMs can call tools”

What people say: “The LLM calls the API” or “The LLM calls the tool”

The reality: An LLM is fundamentally a text-in, text-out system. It generates one token at a time based on probability distributions. It cannot “call” anything. What actually happens in AI agents is a multi-step process:

1. Before inference: Tool descriptions are injected into the prompt, teaching the model what tools exist and their schemas
2. During inference: The LLM is called with a grammar file. It is a file that define what can be present in the final output. During inference the grammar will be checked for each token to prevent constructing an answer that would not be valid. (if you want some lecture you can look at this link)
3. After inference: The agent parses the LLM’s text output, extracts the tool call parameters, and proceed with calling the tool and then preparing step of “thinking” with the result of the tool.

The LLM is producing structured text that describes what should be called. The actual execution happens in the code calling the LLM.

Illustration made using MermaidJS

---

“RAG requires a vector database”

What people say: “To implement RAG, you need to set up a vector store”

The reality: Retrieval-Augmented Generation is an architectural pattern, not a technology requirement. The core concept is simple: instead of asking an LLM a question directly, you first retrieve relevant context, then ask the LLM to answer using that context.

Vector databases are just one option for the retrieval step — and not always the best. You can implement perfectly valid RAG with:

• Traditional full-text search (with Elasticsearch, Solr, …)
• Hybrid search combining keyword and semantic approaches
• Graph databases for relationship-based retrieval
• Simple keyword matching or grep
• …

The “retrieval” in RAG just means “get relevant information somehow.”

---

“LLMs are provided with embeddings during RAG”

What people say: “We embed the documents and feed those embeddings to the model”

The reality: Embeddings are used during the retrieval phase to find semantically similar content, but what gets fed to the LLM is still plain text (or images for multimodal models).

Here’s the actual flow:

1. Query → converted to embedding
2. Embedding → used to search for similar document embeddings
3. Similar documents → retrieved as text
4. Text → inserted into the LLM’s context window
5. LLM → processes text and generates response

The LLM never sees the embeddings. It can’t. The model expects natural language (or image data), not dense vector representations. The embedding space is purely for the retrieval mechanism — it’s how you find the right needles in the haystack, but you still hand the needles (as text) to the LLM, not the magnetic field you used to locate them.

Illustration made using MermaidJS

But please…

know the acronyms you’re using

A few month ago I read a post on LinkedIn on someone telling MCP where know available on ChatGPT, which was right. What wasn’t right it that this person that was promoting himself as an expert was telling MCP was for “Multi-Cloud Provider”

The reality: MCP stands for Model Context Protocol, a standardization effort by Anthropic for connecting LLMs to data sources and tools. It has nothing to do with cloud providers.

It was not the first time I encountered someone making mistakes with acronyms but this one did made a really (bad) impression on me.

---

check the chronology

No, RAG wasn’t created (2020) after the release of ChatGPT (November 30, 2022).

No, Microsoft work on 1.58 Bits quantised LLM wasn’t done on early 2025 (French article) but on early 2024.

---

Conclusion

This list is non-exhaustive and might grow in the future.

As for readers who are genuinely interested in this subject (AI), I urge you to always compare several sources of information (even for this article!). And social media are not valid sources of information!