The AI with an Elephant’s Memory and a Chicken’s Mouth – It Speaks Only What We’ve Taught It

💊 Interactive Test: How Permissive Are AIs with Medical Advice

🧠 How We Built a “Lobotomized” RAG System Using Qdrant, OpenAI, and AI Engine

📌 Purpose:
This document details the functioning of a Retrieval-Augmented Generation (RAG) system designed to provide strictly controlled responses based exclusively on its own internal database, while excluding the general pre-trained knowledge of a language model (LLM).
This process is referred to as “lobotomization.”
The system’s sole source of knowledge consists of four files: one drug catalog (Excel) and three structured medical guidelines (Word documents).

1. 🔍 What is Qdrant?

Qdrant (pronounced Quadrant) is a vector database specialized in storing, managing, and rapidly searching vector embeddings.

Main purpose: It is used in artificial intelligence applications, such as:

🔎 Semantic search

🧩 RAG systems (Retrieval-Augmented Generation)

💬 Specialized chatbots – just like in the use case with the drug database.

2. ⚙️ How Qdrant Works (Conceptual)

The process can be divided into three fundamental stages:

🧩 Conceptual Conclusion: Qdrant does not look for exact text matches; it searches for semantic similarities between meanings.

3. 🏗️ Qdrant Architecture and Usage Modes

Qdrant can be deployed in two main ways:

⚠️ Important: Both options use the same REST or gRPC API, meaning the application code will be identical regardless of the deployment choice.

4. 📚 Data Structure in Qdrant

Qdrant’s data structure is hierarchical and intuitive:

Collection → "medications"
       |
       ├── Point 1 -> Vector + Payload for "Paracetamol"
       ├── Point 2 -> Vector + Payload for "Ibuprofen"
       └── Point n -> Vector + Payload for another medication

Glossary of Terms:

5. 🧮 Query Types in Qdrant

Qdrant supports several query operations, the most relevant of which are:

• 🔍 Simple Search: Returns the most similar vectors for a given query.

• 🧩 Filtered Search: Combines semantic search with metadata filters (e.g., “find antidepressants only from manufacturer Roche”).

• 🔁 Recommend: Generates recommendations based on a set of reference points already in the system.

6. ⚡ Performance and Key Advantages

🧩 Contextual Application: Lobotomized Medication System

🎯 Final Goal: Create a chatbot that responds only based on the 4 files and the 3 text documents, without relying on the general knowledge of the LLM.

🔧 System Components:

1. 🗃️ Vector Database (Qdrant): Stores all information from your files as vectors and payloads.

2. 🛠️ Embedding Engine (OpenAI): Converts user queries and file content into vectors.

3. 🧠 Response Engine (OpenAI GPT): Generates the final textual answer.

4. 🔗 Orchestrator (AI Engine – WordPress Plugin): Connects all components and applies strict rules.

7. 🔐 The Crucial Role of “Severe Instructions” (System Prompt)

This is the soul of the system and the lobotomization mechanism. These instructions are configured in the AI Engine interface and sent to OpenAI GPT at each interaction.

📍 Where do these rules operate?

• NOT in Qdrant.

• NOT in the embedding engine.

• YES, in the response engine (OpenAI GPT), which is required to follow the instructions provided via AI Engine.

📜 Example of Severe Instructions for AI Engine:

# IDENTITY: You are a specialized assistant for the company's medication database.
# KNOWLEDGE SOURCE: Your only source of information is the vector database (Qdrant), which contains only files X, Y, Z.
# ABSOLUTE RULES:
# 1. Respond EXCLUSIVELY based on text fragments (context) returned by Qdrant for the user's question.
# 2. NEVER use general pre-trained knowledge about medications, diseases, or any other topic.
# 3. IF the fragments returned by Qdrant contain no relevant information for the question, your answer must be STRICTLY: "No information available on this topic in our database."
# 4. Answers must be concise, clear, and based solely on the provided data.

8. 🔁 Data Flow and Lobotomization in Action (Extended Examples)

Let’s follow a few concrete scenarios to understand how lobotomization and smart information retrieval work.

Use Case 1: Searching for a Specific Medication (Strict Knowledge Control)

Use Case 2: From Symptom to Diagnosis and Treatment (Full Knowledge Flow)

Use Case 3: Advanced Semantic Search (Without Exact Keywords)

Use Case 4: Implicit Filtering via File Content

Use Case 5: Excluding Any External Information

9. 🧠 How Logical Connections Between Files Are Made

• Who makes the connections? Not Qdrant, but OpenAI GPT.

• How?

  1. All files (Excel and DOCs) are “chunked” into smaller text pieces.
  2. Each chunk is vectorized and inserted into Qdrant as an independent point.
  3. Upon a query, Qdrant returns the most relevant fragments, regardless of which file they come from.
  4. GPT receives these mixed fragments and synthesizes them, finding semantic links between them to build a coherent answer.

Example: If one fragment mentions “Paracetamol” and another “liver contraindications,” and the user asks about liver effects, GPT connects the two fragments to provide a complete answer.

🔗 System Knowledge Structure: The vector database is built on a set of 4 files forming a complete clinical hierarchy:

1. Symptoms.doc – Structured list of symptoms (e.g., Headache, Vertigo, Muscle Weakness).
2. Diagnoses.doc – Possible diagnoses for each symptom.
3. Treatments.doc – Recommended treatments for each diagnosis.
4. Excel.xlsx – Catalog of specific commercial products (Name, INN, Manufacturer, ATC Code, etc.).

This structure allows the system to follow a complete logical chain: from the patient’s symptom to the appropriate diagnoses, to the recommended treatment, and finally to the specific medications available in the catalog.

✅ Final Conclusion

The system built is robust and perfectly achieves the goal of providing a specialized and controlled chatbot:

• Qdrant is the high-performance semantic memory.
• OpenAI is the brain that understands and generates text.
• AI Engine is the conductor that applies the score (severe instructions).
• Severe Instructions are the logical quarantine ensuring lobotomization, forcing the system to respond only from the specific data source.

This architecture eliminates the risk of “hallucinations” (invented responses) and provides full control over information, making it ideal for sensitive domains such as medical applications. The final system will function as a virtual neurology expert, reporting exclusively on the knowledge base structured in the 4 files, providing safe, traceable, and verified answers.