Skip to content

cheat_sheet.md

Latest commit

 

History

History
264 lines (213 loc) · 8.49 KB

cheat_sheet.md

File metadata and controls

264 lines (213 loc) · 8.49 KB

🔥 Generative AI – Complete Cheat Sheet

One page with everything you need. Keep this open in a second window.

Core Concepts (One Line Each)

Concept What it is
Token ~4 chars of text. The atom of LLMs. Costs money.
Prompt Everything fed to the model: system + history + user message
Temperature How random the output is. 0=deterministic, 1=creative
Embedding Text converted to a vector. Similar text = similar vectors
RAG Give the AI relevant document chunks right before it answers
Fine-tuning Retrain on your data to change behavior/style. Use LoRA.
Function calling AI outputs a tool command; your code executes it
Agent LLM in a loop; takes multiple actions to complete a task
Hallucination Model confidently says something false. Prevent with RAG.
Context window How much text the model can "see" at once. Measured in tokens.
System prompt Instructions given before the conversation. Sets behavior.
Few-shot Giving examples in the prompt to guide format/style
Chain-of-thought Ask model to reason step-by-step before answering
Vector DB Database for storing embeddings. Enables fast similarity search.

Hugging Face Snippets

Basic setup (always start with this)

from huggingface_hub import InferenceClient
from dotenv import load_dotenv
import os

load_dotenv()
client = InferenceClient(token=os.getenv("HUGGINGFACEHUB_API_TOKEN"))

Chat completion (streaming)

def chat(messages, model="deepseek-ai/DeepSeek-R1", temp=0.7, max_tokens=500):
    response = client.chat_completion(
        messages=messages, model=model,
        temperature=temp, max_tokens=max_tokens, stream=True
    )
    full = ""
    for chunk in response:
        if hasattr(chunk, 'choices') and chunk.choices:
            content = chunk.choices[0].delta.content
            if content:
                full += content
                print(content, end="", flush=True)
    print()
    return full

Chat completion (non-streaming, simpler)

response = client.chat_completion(
    messages=[{"role": "user", "content": "Hello"}],
    model="deepseek-ai/DeepSeek-R1",
    max_tokens=200,
    stream=False
)
text = response.choices[0].message.content

Messages format

messages = [
    {"role": "system", "content": "You are a helpful assistant."},  # optional
    {"role": "user", "content": "What is Python?"},
    {"role": "assistant", "content": "Python is a programming language."},  # history
    {"role": "user", "content": "Is it good for beginners?"},  # current
]

Local embeddings

from sentence_transformers import SentenceTransformer, util

model = SentenceTransformer('all-MiniLM-L6-v2')
embeddings = model.encode(["text1", "text2"])
similarity = util.cos_sim(embeddings[0], embeddings[1]).item()  # 0.0 to 1.0

RAG in 10 lines

from sentence_transformers import SentenceTransformer, util
import numpy as np

embed_model = SentenceTransformer('all-MiniLM-L6-v2')
chunks = ["chunk1 text", "chunk2 text", "chunk3 text"]
chunk_embeddings = embed_model.encode(chunks)

question = "your question"
q_embedding = embed_model.encode(question)
similarities = util.cos_sim(q_embedding, chunk_embeddings)[0]
best_idx = similarities.argmax().item()
best_chunk = chunks[best_idx]
# Now put best_chunk in your prompt as context

ChromaDB (persistent vector DB)

import chromadb
from sentence_transformers import SentenceTransformer

client_db = chromadb.Client()
collection = client_db.create_collection("my_docs")
embedder = SentenceTransformer('all-MiniLM-L6-v2')

# Add documents
texts = ["doc1", "doc2", "doc3"]
embeddings = embedder.encode(texts).tolist()
collection.add(documents=texts, embeddings=embeddings, ids=["id1", "id2", "id3"])

# Query
q_emb = embedder.encode(["my question"]).tolist()
results = collection.query(query_embeddings=q_emb, n_results=3)
relevant_docs = results['documents'][0]

Image captioning

with open("photo.jpg", "rb") as f:
    caption = client.image_to_text(f.read(), model="Salesforce/blip-image-captioning-large")
print(caption)

Text-to-image

image = client.text_to_image("a cat on a sofa", model="black-forest-labs/FLUX.1-dev")
image.save("cat.png")  # PIL Image object

Visual Q&A

with open("photo.jpg", "rb") as f:
    result = client.visual_question_answering(
        f.read(), question="What color is it?", model="dandelin/vilt-b32-finetuned-vqa"
    )

Free Models Quick Reference

Task Model ID Notes
Best general chat deepseek-ai/DeepSeek-R1 Reasoning model, very capable
Fast chat meta-llama/Llama-3.1-8B-Instruct Quick, good quality
General chat HuggingFaceH4/zephyr-7b-beta Reliable free-tier model
Multilingual Qwen/Qwen2.5-7B-Instruct Good for non-English
Local embeddings sentence-transformers/all-MiniLM-L6-v2 Fast, 384-dim
Better embeddings sentence-transformers/all-mpnet-base-v2 Slower, 768-dim, more accurate
Image captioning Salesforce/blip-image-captioning-large Best free captioner
Visual Q&A dandelin/vilt-b32-finetuned-vqa Simple visual questions
Text-to-image black-forest-labs/FLUX.1-dev Highest quality free
Text-to-image (alt) stabilityai/stable-diffusion-2-1 Alternative if FLUX unavailable
Vision LLM meta-llama/Llama-3.2-11B-Vision-Instruct Complex image understanding
Code bigcode/starcoder2-15b Code generation
Fine-tuning base microsoft/phi-2 Small, fine-tunable on Colab

Temperature Guide

Temperature Use for
0.0 Code generation, structured output, JSON extraction
0.1–0.3 Factual Q&A, classification, evaluation
0.4–0.6 General chat, explanations, summaries
0.7–0.8 Creative writing, brainstorming
0.9–1.0 Poetry, experimental, maximum variety

Prompt Templates

JSON extraction

Extract [WHAT] from the text below.
Return ONLY a JSON object: {"key1": ..., "key2": ...}
If a field is missing, use null. No explanation, no markdown.

Text: [INPUT]

Code review

You are a senior [LANGUAGE] engineer. Review this code.
For each issue: Issue | Severity (Critical/High/Medium/Low) | Fix
Code: [CODE]

Summarization

Summarize the following [CONTENT TYPE] for [AUDIENCE].
Length: [SHORT/MEDIUM/DETAILED]. Format: [BULLETS/PARAGRAPH]
[CONTENT]

RAG answer

Answer the question based ONLY on the provided context.
If not in context, say "I couldn't find that in the document."
Context: [RETRIEVED CHUNKS]
Question: [USER QUESTION]

Error Quick Reference

Error Fix
Model not supported by any provider Switch to deepseek-ai/DeepSeek-R1
Rate limit exceeded Add retry with time.sleep(60)
404 Not Found Use chat_completion not text_generation
Invalid token Check .env file, token starts with hf_
json.JSONDecodeError Strip thinking blocks: re.sub(r'<think>.*?</think>', '', text)
IndexError on chunk.choices Add if hasattr(chunk, 'choices') and chunk.choices: check
CUDA out of memory Reduce batch size, use 4-bit quantization (QLoRA)

Project Files Summary

File Command What it does
projects/prompt_playground.py streamlit run ... A/B test prompts with temperature control
projects/rag_pdf_chat.py streamlit run ... Chat with any PDF
projects/agent.py streamlit run ... Multi-tool AI agent
projects/evaluate.py streamlit run ... Score and compare AI outputs
projects/multimodal_app.py streamlit run ... Image captioning, VQA, text-to-image

Decision Guide

Do you need the AI to know YOUR specific documents/data?
  → YES → RAG (03_rag.md)
  
Do you need the AI to take actions in the real world?
  → YES → Function Calling (04a, 04b)
  
Do you need to change the AI's personality/style/format consistently?
  → YES → Fine-tuning (05a, 05b)
  
Is the AI's output bad/unreliable?
  → Need better quality → Improve system prompt or use RAG
  → Need to measure how bad → Evaluation (06_evaluation.md)
  
Do you need to understand images?
  → YES → Multi-modal (07_multimodal.md)
  
Want to deploy publicly?
  → YES → Production (10_production.md)

Happy building! 🚀