---
type: Note
---
# PROMPT: generowanie projektu Generative AI (LLM + RAG)

Jasne — przygotuję Ci **gotowy, produkcyjny prompt**, który możesz wrzucić do **Claude Code / Copilot / dowolnego agenta**, żeby **automatycznie wygenerował projekt** w tej architekturze.

Dam Ci wersję:

- ✅ deterministyczną (minimalne „halucynacje”)
- ✅ zgodną z Twoim stylem (DBA + SOP + operacyjność)
- ✅ gotową do użycia w CLI (`claude -p`)

***

# ✅ PROMPT: generowanie projektu Generative AI (LLM + RAG)

```text
You are a senior software architect and platform engineer.

Your task is to generate a complete, production-ready Generative AI project based on the following requirements.

# CONTEXT
The project must implement:
- LLM abstraction (support OpenAI, Claude, local models)
- Retrieval-Augmented Generation (RAG)
- Data processing pipeline (chunking, tokenization, preprocessing)
- Prompt management system
- Inference engine
- Testing and automation scripts
- Container-ready (Docker)

The system must be structured, maintainable, and suitable for enterprise environments.

# REQUIREMENTS

## 1. Project structure
Create a full file tree exactly in this format:

generative_ai_project/
  config/
  data/
  src/
    core/
    prompts/
    rag/
    processing/
    inference/
  tests/
    unit/
    integration/
  scripts/

Include all files and implement real code (not placeholders).

---

## 2. Core architecture

### src/core/
Implement:
- base_llm.py → abstract interface (send, generate, stream)
- gpt_client.py → OpenAI client
- claude_client.py → Claude client
- local_llm.py → local model stub
- model_factory.py → dynamic model selection based on config

Requirements:
- retries
- timeout handling
- structured logging

---

## 3. Prompt layer (src/prompts/)
- templates.py → prompt templates (Jinja-like)
- chain.py → multi-step chaining

Requirements:
- reusable prompt structure
- system + user + context separation

---

## 4. RAG module (src/rag/)
Implement:
- embedder.py → embedding generation
- retriever.py → similarity search (top-k)
- vector_store.py → wrapper (FAISS or simple in-memory version)
- indexer.py → indexing pipeline

Requirements:
- deterministic chunking
- metadata support
- clear interfaces between components

---

## 5. Data processing (src/processing/)
Implement:
- chunking.py → configurable chunk size + overlap
- tokenizer.py → basic tokenizer abstraction
- preprocessor.py → cleaning pipeline

---

## 6. Inference (src/inference/)
Implement:
- inference_engine.py → orchestration:
  user_query → retriever → prompt → LLM → response parser
- response_parser.py → structured output (JSON-safe)

---

## 7. Config (config/)
Create:
- model_config.yaml
- logging_config.yaml

Requirements:
- environment-aware (dev/prod)
- no hardcoded secrets

---

## 8. Scripts (scripts/)
Create:
- setup_env.sh
- run_tests.sh
- build_embeddings.py
- cleanup.py

---

## 9. Tests (tests/)
Create:
- unit tests for LLM clients and prompts
- integration test for full RAG + inference flow

---

## 10. Operational requirements
All code must:
- include logging
- include error handling
- support debugging
- avoid global state

---

## 11. Output format
Return:
1. Full directory tree
2. Full code for each file
3. Example config files
4. Example usage (CLI or script)

Do not skip files. Do not leave TODOs.

---

# CONSTRAINTS
- Code must be Python
- Clean architecture (separation of concerns)
- No unnecessary dependencies
- Ready to run locally

---

# BONUS
If possible:
- add simple CLI entrypoint
- add example dataset
- include README structure

# GOAL
The result must be a working skeleton that can be extended into production.

Start now.
```

***

# ✅ Jak tego użyć (praktycznie)

## Claude Code (najlepsza opcja)

```bash
claude --bare -p "$(cat prompt.txt)"
```

albo:

```bash
claude --bare -p "<<<tu wklej prompt>>>"
```

***

## 🔥 PRO TIP (dla Ciebie – advanced)

Jeśli chcesz mieć **100% powtarzalności (bardzo w Twoim stylu)**:

Dodaj do prompta:

```text
Use deterministic behavior:
- avoid random examples
- use fixed sample data
- keep outputs stable
```

***

# ✅ Wersja rozszerzona (dla Twojego systemu – OpenClaw / agents)

Możesz to też użyć jako **skill/mission prompt**:

```text
MISSION: generate-genai-project

INPUT:
- project_name
- llm_provider (gpt/claude/local)
- use_rag (true/false)

OUTPUT:
- full repo structure
- working code
- scripts + tests

CONSTRAINTS:
- deterministic output
- no placeholders
- production-ready
```

***

# Jeśli chcesz dalej

Mogę Ci zrobić kolejne poziomy:

✅ prompt do:

- generowania **tylko RAG pipeline**
- generowania **tylko inference engine**
- generowania **runbooków + monitoring + alerts**

✅ albo:

- wersję pod **Twoje realne środowisko (Proxmox / on-prem / DB-backed RAG)**
- integrację z **Splunk / ServiceNow / Agentami (MCC/APOLLO)**

Powiedz kierunek → zrobię Ci to już pod konkretne wdrożenie 👍