Prompt Engineering & Context Engineering with LangChain and OpenAI

Scripts demonstrating 45 Prompt Engineering, Context Engineering, and Agentic AI techniques using LangChain and the OpenAI API.

Language / Idioma: Português Brasileiro | English

Study Roadmap: Want to master AI-assisted development? Check out the ROADMAP.md - a comprehensive guide with learning tracks (beginner to advanced), technique connection maps, project organization templates, and modern tooling integration with Claude Code.

Implemented Techniques

Prompt Engineering - Basic (01-06)

Script	Technique	Description
01_zero_shot.py	Zero-Shot	Direct prompts without prior examples
02_chain_of_thought.py	Chain of Thought (CoT)	Step-by-step reasoning
03_few_shot.py	Few-Shot	Examples to guide the model
04_tree_of_thoughts.py	Tree of Thoughts (ToT)	Multiple reasoning paths
05_skeleton_of_thought.py	Skeleton of Thought (SoT)	Structure first, details later
06_react_agent.py	ReAct	Reasoning + Actions with tools

Prompt Engineering - Advanced (07-10)

Script	Technique	Description
07_self_consistency.py	Self-Consistency	Generate N responses, vote on most consistent
08_least_to_most.py	Least-to-Most	Progressive decomposition into sub-problems
09_self_refine.py	Self-Refine	Iterative critique and improvement
10_prompt_chaining.py	Prompt Chaining	Pipeline of connected prompts

Context Engineering - RAG (11-13)

Script	Technique	Description
11_rag_basic.py	RAG Basic	ChromaDB + semantic search + chunking
12_rag_reranking.py	RAG + Reranking	Reordering for better relevance
13_rag_conversational.py	Conversational RAG	RAG with chat memory

Local Models - Ollama (14-15)

Script	Technique	Description
14_ollama_basic.py	Ollama Basic	Local LLMs (Llama 3, Mistral)
15_ollama_rag.py	Ollama + RAG	100% offline RAG

Structured Output & Tools (16-17)

Script	Technique	Description
16_structured_output.py	Structured Output	JSON mode + Pydantic models
17_tool_calling.py	Tool Calling	Custom function tools

Advanced Features (18-20)

Script	Technique	Description
18_vision_multimodal.py	Vision/Multimodal	Image analysis with GPT-4o
19_memory_conversation.py	Memory/Conversation	Persistent conversation context
20_meta_prompting.py	Meta-Prompting	LLM generating/optimizing prompts

Context Engineering - Chunking & Retrieval (21-25)

Script	Technique	Description
21_advanced_chunking.py	Advanced Chunking	Semantic, recursive, token-based, sliding window strategies
22_hybrid_search.py	Hybrid Search	BM25 (keyword) + Vector (semantic) with RRF fusion
23_query_transformation.py	Query Transformation	HyDE, Multi-Query, Step-Back, Decomposition
24_contextual_compression.py	Contextual Compression	Extract only relevant parts from documents
25_self_query.py	Self-Query Retrieval	LLM auto-generates metadata filters

Context Engineering - Context Management (26-30)

Script	Technique	Description
26_parent_document.py	Parent-Document Retrieval	Small chunks for search, large parents for context
27_multi_vector.py	Multi-Vector Retrieval	Multiple representations (summary + questions + content)
28_ensemble_retrieval.py	Ensemble Retrieval	Combine multiple retrievers with weighted RRF
29_long_context.py	Long Context Strategies	Map-Reduce, Refine, Map-Rerank for large documents
30_time_weighted.py	Time-Weighted Retrieval	Recency bias in retrieval with exponential decay

MCP & Agentic AI (31-35)

Script	Technique	Description
31_mcp_basics.py	MCP Basics	Model Context Protocol fundamentals (resources, tools, prompts)
32_mcp_server_stdio.py	MCP Server STDIO	Local MCP server with standard input/output transport
33_mcp_server_http.py	MCP Server HTTP/SSE	Remote MCP server with HTTP and Server-Sent Events
34_multi_agent.py	Multi-Agent	Collaborative AI agents (pipeline, debate, hierarchical patterns)
35_prompt_evaluation.py	Prompt Evaluation	Evaluate prompt quality, A/B testing, observability

Enterprise & Production (36-40)

Script	Technique	Description
36_llm_security.py	LLM Security	OWASP Top 10, prompt injection detection, guardrails, rate limiting
37_caching_strategies.py	Caching Strategies	Response cache, semantic cache, embedding cache, conversation cache
38_cost_optimization.py	Cost Optimization	Token counting, model selection, usage tracking, budget management
39_ai_testing.py	AI Testing	Testing non-deterministic outputs, validators, mocking, snapshots
40_fine_tuning.py	Fine-tuning	Dataset preparation, validation, fine-tuning workflow, best practices

Agentic AI - Advanced (41-45)

Script	Technique	Description
41_agent_skills.py	Agent Skills	On-demand knowledge loading, SEO-like skill descriptions, lazy loading
42_context_window.py	Context Window Management	Monitor context usage, smart summarization, context health tracking
43_subagent_orchestration.py	Subagent Orchestration	Isolated context windows, parallel execution, results aggregation
44_shared_memory.py	Shared Memory	Memory DB for agents, short/medium/long-term memory, feedback loops
45_spec_generation.py	Spec-Driven Development	Technical specification generation, validation, task breakdown

Requirements

• Python 3.10+
• OpenAI API key
• (Optional) Ollama for local models
• (Optional) Cohere API key for reranking

Installation

1. Clone or navigate to the project directory:

cd /path/to/project

2. Create and activate a virtual environment:

python -m venv venv
source venv/bin/activate  # Linux/macOS
# or
venv\Scripts\activate     # Windows

3. Install dependencies:

pip install -r requirements.txt

4. Configure credentials:

cp .env.example .env

Edit the .env file and add your keys:

OPENAI_API_KEY=sk-your-key-here
OPENAI_MODEL=gpt-4o-mini

# Optional - for Ollama (local models)
OLLAMA_MODEL=llama3.2
OLLAMA_BASE_URL=http://localhost:11434

# Optional - for Cohere reranking
COHERE_API_KEY=your-cohere-key-here

5. (Optional) Install Ollama for local models:

# macOS
brew install ollama

# Linux
curl -fsSL https://ollama.com/install.sh | sh

# Start Ollama service
ollama serve

# Pull a model
ollama pull llama3.2
ollama pull nomic-embed-text  # For embeddings

Usage

Execute any script from the techniques/ folder:

English examples:

# Basic Prompting (01-06)
python techniques/en/01_zero_shot.py
python techniques/en/02_chain_of_thought.py
python techniques/en/03_few_shot.py
python techniques/en/04_tree_of_thoughts.py
python techniques/en/05_skeleton_of_thought.py
python techniques/en/06_react_agent.py

# Advanced Prompting (07-10)
python techniques/en/07_self_consistency.py
python techniques/en/08_least_to_most.py
python techniques/en/09_self_refine.py
python techniques/en/10_prompt_chaining.py

# RAG (11-13) - Requires sample_data/
python techniques/en/11_rag_basic.py
python techniques/en/12_rag_reranking.py
python techniques/en/13_rag_conversational.py

# Ollama (14-15) - Requires Ollama running
python techniques/en/14_ollama_basic.py
python techniques/en/15_ollama_rag.py

# Structured Output & Tools (16-17)
python techniques/en/16_structured_output.py
python techniques/en/17_tool_calling.py

# Advanced Features (18-20)
python techniques/en/18_vision_multimodal.py
python techniques/en/19_memory_conversation.py
python techniques/en/20_meta_prompting.py

# Context Engineering - Chunking & Retrieval (21-25)
python techniques/en/21_advanced_chunking.py
python techniques/en/22_hybrid_search.py
python techniques/en/23_query_transformation.py
python techniques/en/24_contextual_compression.py
python techniques/en/25_self_query.py

# Context Engineering - Context Management (26-30)
python techniques/en/26_parent_document.py
python techniques/en/27_multi_vector.py
python techniques/en/28_ensemble_retrieval.py
python techniques/en/29_long_context.py
python techniques/en/30_time_weighted.py

# MCP & Agentic AI (31-35)
python techniques/en/31_mcp_basics.py
python techniques/en/32_mcp_server_stdio.py
python techniques/en/33_mcp_server_http.py
python techniques/en/34_multi_agent.py
python techniques/en/35_prompt_evaluation.py

# Enterprise & Production (36-40)
python techniques/en/36_llm_security.py
python techniques/en/37_caching_strategies.py
python techniques/en/38_cost_optimization.py
python techniques/en/39_ai_testing.py
python techniques/en/40_fine_tuning.py

# Agentic AI - Advanced (41-45)
python techniques/en/41_agent_skills.py
python techniques/en/42_context_window.py
python techniques/en/43_subagent_orchestration.py
python techniques/en/44_shared_memory.py
python techniques/en/45_spec_generation.py

Portuguese examples:

python techniques/pt-br/01_zero_shot.py
# ... (same pattern with pt-br/)

Technique Descriptions

1. Zero-Shot Prompting

Technique where the model receives a task without prior examples, using only its pre-trained knowledge.

Available functions:

• classify_sentiment(text) - Classifies sentiment as POSITIVE, NEGATIVE, or NEUTRAL
• translate_text(text, target_language) - Translates text to specified language
• extract_entities(text) - Extracts people, locations, organizations, and dates
• summarize_text(text) - Summarizes text in a few sentences

---

2. Chain of Thought (CoT)

Instructs the model to "think step by step" before reaching the final answer, improving performance on reasoning tasks.

Available functions:

• solve_math_problem(problem) - Solves math problems showing each step
• logical_reasoning(puzzle) - Solves logic puzzles with deductions
• analyze_decision(situation) - Analyzes scenarios for decision making
• debug_code(code, error) - Analyzes code and error to find solution

---

3. Few-Shot Prompting

Provides examples to the model before the task, helping it understand the format and expected response type.

Available functions:

• classify_support_ticket(ticket) - Classifies tickets with category, priority, and action
• convert_to_sql(description) - Converts natural language to SQL
• generate_docstring(code) - Generates Google Style docstrings
• extract_structured_data(text) - Extracts data in JSON format

---

4. Tree of Thoughts (ToT)

Explores multiple reasoning paths in parallel, evaluates each one, and selects the most promising.

Available functions:

• tree_of_thoughts(problem, depth) - Executes complete ToT algorithm
• generate_thoughts(problem, num) - Generates multiple initial approaches
• evaluate_thought(problem, thought) - Evaluates viability of an approach
• expand_thought(problem, thought, next_step) - Develops an approach

---

5. Skeleton of Thought (SoT)

First generates a "skeleton" (structure/topics) and then expands each part, allowing parallelization.

Available functions:

• skeleton_of_thought_sync(topic, context) - Synchronous version
• skeleton_of_thought_async(topic, context) - Asynchronous version (parallel)
• generate_skeleton(topic, context) - Generates list of topics
• expand_topic(main_topic, topic, context) - Expands a specific topic

---

6. ReAct Agent

Combines reasoning (Thought) with actions (Action) and observations (Observation) in an iterative loop, using external tools.

Available tools:

• web_search - Internet search via DuckDuckGo
• wikipedia - Wikipedia queries
• calculator - Mathematical calculations

---

7. Self-Consistency

Generates multiple responses to the same problem, then uses majority voting to select the most consistent answer. Improves accuracy on reasoning tasks.

Available functions:

• self_consistency_solve(problem, num_samples) - Solves with multiple samples and voting
• solve_with_voting(problem, num_samples) - Alternative with explicit voting
• extract_answer(response) - Extracts final answer from response

Example:

from techniques.en.self_consistency import self_consistency_solve

result = self_consistency_solve(
    "If a train travels 120 km in 2 hours, what is its speed?",
    num_samples=5
)
print(result["final_answer"])  # 60 km/h

---

8. Least-to-Most Prompting

Decomposes complex problems into smaller sub-problems, solves them progressively from simplest to most complex, building on previous answers.

Available functions:

• least_to_most_solve(problem) - Complete decomposition and solution
• decompose_problem(problem) - Breaks down into sub-problems
• solve_subproblem(subproblem, context) - Solves with previous context

Example:

from techniques.en.least_to_most import least_to_most_solve

result = least_to_most_solve(
    "How do I build a machine learning model to predict house prices?"
)
print(result["final_answer"])

---

9. Self-Refine

Generates an initial response, then iteratively critiques and improves it until it meets quality standards.

Available functions:

• self_refine(task, max_iterations) - Complete refinement loop
• generate_initial(task) - Creates first draft
• critique(task, response) - Evaluates and identifies issues
• refine(task, response, feedback) - Improves based on critique

Example:

from techniques.en.self_refine import self_refine

result = self_refine(
    "Write a function to check if a string is a palindrome",
    max_iterations=3
)
print(result["final_response"])

---

10. Prompt Chaining

Connects multiple prompts in a pipeline where the output of one becomes the input of the next, enabling complex multi-step workflows.

Available functions:

• chain_prompts(initial_input, prompt_chain) - Executes prompt pipeline
• research_chain(topic) - Research → Analysis → Summary
• content_chain(topic) - Outline → Draft → Edit → Format

Example:

from techniques.en.prompt_chaining import content_chain

result = content_chain("Benefits of Remote Work")
print(result["final_output"])

---

11. RAG Basic

Retrieval-Augmented Generation with ChromaDB for document storage, semantic search, and text chunking.

Available functions:

• create_vectorstore(documents) - Creates ChromaDB vector store
• rag_query(question, vectorstore) - Queries with RAG
• load_and_split_documents(path) - Loads and chunks documents

Key features:

• Recursive text chunking (1000 chars, 200 overlap)
• OpenAI embeddings for semantic search
• Top-k retrieval with relevance scores

Example:

from techniques.en.rag_basic import create_vectorstore, rag_query

# Load documents and create vector store
vectorstore = create_vectorstore(documents)

# Query with RAG
result = rag_query("What is machine learning?", vectorstore)
print(result["answer"])

---

12. RAG + Reranking

Enhances basic RAG with reranking to improve retrieval relevance. Supports multiple reranking methods.

Reranking methods:

• LLM-based reranking (uses GPT to score relevance)
• Cohere Rerank (requires API key)
• CrossEncoder (local transformer model)

Available functions:

• rag_with_reranking(question, vectorstore, method) - RAG with reranking
• llm_rerank(question, documents) - LLM-based reranking
• cohere_rerank(question, documents) - Cohere API reranking

---

13. Conversational RAG

RAG with conversation memory for multi-turn dialogues. Maintains context across questions.

Memory types:

• Buffer Memory - Full conversation history
• Summary Memory - Compressed summary

Available functions:

• create_conversational_rag(vectorstore) - Creates conversational chain
• chat(question) - Chat with memory
• get_chat_history() - Retrieve conversation history

---

14. Ollama Basic

Use local LLMs via Ollama without API costs or internet dependency.

Supported models:

• llama3.2 - Meta's Llama 3
• mistral - Mistral 7B
• codellama - Code-specialized Llama
• phi3 - Microsoft's Phi-3

Available functions:

• ollama_chat(message) - Chat with local model
• ollama_generate(prompt) - Text generation
• list_local_models() - List available models

Example:

from techniques.en.ollama_basic import ollama_chat

response = ollama_chat("Explain quantum computing in simple terms")
print(response)

---

15. Ollama + RAG

100% offline RAG using Ollama for both embeddings and generation.

Components:

• Local embeddings: nomic-embed-text
• Local LLM: llama3.2 or mistral
• ChromaDB for vector storage

Available functions:

• create_local_vectorstore(documents) - Creates store with local embeddings
• local_rag_query(question, vectorstore) - Query with local RAG

---

16. Structured Output

Force LLM outputs to follow specific schemas using Pydantic models or JSON mode.

Available functions:

• extract_person(text) - Extract person info as Pydantic model
• extract_invoice(text) - Extract invoice data
• json_mode_extract(text, schema) - Generic JSON extraction

Example:

from techniques.en.structured_output import extract_person
from pydantic import BaseModel

class Person(BaseModel):
    name: str
    age: int
    occupation: str

result = extract_person("John is a 30-year-old software engineer")
print(result.name)  # John
print(result.age)   # 30

---

17. Tool Calling

Enable LLMs to call custom functions/tools to perform actions or retrieve information.

Available tools:

• get_weather(city) - Get weather information
• calculate(expression) - Perform calculations
• search_database(query) - Search mock database

Example:

from techniques.en.tool_calling import agent_with_tools

response = agent_with_tools(
    "What's the weather in Tokyo and calculate 15% tip on $85"
)
print(response)

---

18. Vision/Multimodal

Analyze images using vision-enabled models like GPT-4o.

Available functions:

• analyze_image(image_path, prompt) - Analyze image with custom prompt
• describe_image(image_path) - Generate detailed description
• extract_text_from_image(image_path) - OCR-like text extraction
• analyze_chart(image_path) - Analyze charts and graphs
• compare_images(image1, image2) - Compare two images

Example:

from techniques.en.vision_multimodal import analyze_chart

result = analyze_chart("sample_data/images/chart.png")
print(result)  # Chart type, data, insights

---

19. Memory/Conversation

Maintain conversation context across multiple interactions using different memory strategies.

Memory types:

• BufferMemory - Stores complete conversation history
• WindowMemory - Stores last N exchanges
• SummaryMemory - Maintains compressed summary
• EntityMemory - Tracks mentioned entities

Example:

from techniques.en.memory_conversation import ConversationChain

chain = ConversationChain(memory_type="buffer")
response1 = chain.chat("My name is Alice")
response2 = chain.chat("What's my name?")  # Remembers "Alice"

---

20. Meta-Prompting

Use an LLM to generate, optimize, and improve prompts for other LLM tasks.

Available functions:

• generate_prompt(task_description) - Generate optimized prompt
• optimize_prompt(original_prompt, issues) - Improve existing prompt
• evaluate_prompt(prompt, task) - Score and critique a prompt
• generate_prompt_variations(base_prompt) - A/B testing variations
• auto_improve_prompt(prompt, task, test_input) - Iterative improvement

Example:

from techniques.en.meta_prompting import generate_prompt

prompt = generate_prompt(
    task_description="Extract key information from customer emails",
    context="SaaS company support",
    constraints=["JSON output", "Include urgency level"]
)
print(prompt)

---

21. Advanced Chunking

Multiple text splitting strategies optimized for different content types and retrieval scenarios.

Chunking strategies:

• RecursiveCharacter - Hierarchical splitting by separators
• TokenBased - Split by token count (model-aware)
• MarkdownAware - Respects markdown structure
• Semantic - Groups by semantic similarity
• SlidingWindow - Overlapping fixed-size windows
• SentenceBased - Natural sentence boundaries

Available functions:

• recursive_character_chunking(text, chunk_size, overlap) - Standard recursive splitting
• token_based_chunking(text, chunk_size) - Token-aware splitting
• markdown_aware_chunking(text) - Structure-preserving for markdown
• semantic_chunking(text, threshold) - Similarity-based grouping
• sliding_window_chunking(text, window_size, step) - Overlapping windows
• sentence_based_chunking(text, sentences_per_chunk) - Sentence grouping

Example:

from techniques.en.advanced_chunking import semantic_chunking

chunks = semantic_chunking(long_document, threshold=0.75)
for chunk in chunks:
    print(f"Chunk: {len(chunk)} chars")

---

22. Hybrid Search

Combines keyword-based (BM25) and semantic (vector) search using Reciprocal Rank Fusion.

Components:

• BM25Retriever - Traditional keyword matching
• VectorRetriever - Semantic similarity search
• HybridRetriever - Weighted combination

Available functions:

• create_hybrid_retriever(documents, bm25_weight, vector_weight) - Create hybrid retriever
• reciprocal_rank_fusion(results_list, k) - Combine ranked results
• hybrid_search(query, k) - Search with both methods

Example:

from techniques.en.hybrid_search import HybridRetriever

retriever = HybridRetriever(documents, bm25_weight=0.4, vector_weight=0.6)
results = retriever.search("machine learning algorithms", k=5)

---

23. Query Transformation

Transform queries to improve retrieval effectiveness using various techniques.

Transformation methods:

• HyDE - Hypothetical Document Embeddings (generate hypothetical answer, search with that)
• Multi-Query - Generate multiple query variations
• Step-Back - Abstract query to broader concept
• Decomposition - Break complex query into sub-queries

Available functions:

• hyde_transform(query) - Generate hypothetical document
• multi_query_transform(query, num_queries) - Generate query variations
• step_back_transform(query) - Abstract to broader question
• decompose_query(query) - Split into sub-questions

Example:

from techniques.en.query_transformation import multi_query_transform

queries = multi_query_transform(
    "What are the best practices for microservices?",
    num_queries=3
)
# Returns variations like:
# - "microservices architecture best practices"
# - "how to design microservices effectively"
# - "recommended patterns for microservice development"

---

24. Contextual Compression

Extract only the relevant portions of retrieved documents to reduce noise and token usage.

Compression methods:

• LLMExtractor - Use LLM to extract relevant sentences
• EmbeddingsFilter - Filter by semantic similarity
• SentenceExtractor - Extract relevant sentences by scoring

Available functions:

• create_compression_retriever(base_retriever, compressor) - Wrap retriever with compression
• llm_compress(documents, query) - LLM-based compression
• embeddings_filter(documents, query, threshold) - Similarity-based filtering

Example:

from techniques.en.contextual_compression import ContextualCompressionRetriever

compression_retriever = ContextualCompressionRetriever(
    base_retriever=vector_retriever,
    compressor=LLMExtractorCompressor()
)
# Returns only relevant excerpts instead of full documents
results = compression_retriever.retrieve("What is RAG?")

---

25. Self-Query Retrieval

LLM automatically generates metadata filters from natural language queries.

Features:

• Automatic filter extraction from queries
• Support for comparison operators (=, >, <, >=, <=)
• Combines semantic search with structured filtering

Available functions:

• create_self_query_retriever(vectorstore, metadata_info) - Create self-query retriever
• parse_query(query) - Extract semantic query and filters
• apply_filters(documents, filters) - Apply metadata filters

Example:

from techniques.en.self_query import SelfQueryRetriever

retriever = SelfQueryRetriever(
    vectorstore=vectorstore,
    metadata_fields=[
        {"name": "category", "type": "string"},
        {"name": "price", "type": "float"},
        {"name": "year", "type": "integer"}
    ]
)

# Query: "cheap electronics from 2024"
# Auto-generates: category="electronics", price<100, year=2024
results = retriever.retrieve("cheap electronics from 2024")

---

26. Parent-Document Retrieval

Search with small chunks for precision, but retrieve larger parent documents for context.

Concept:

• Child chunks: Small (e.g., 400 chars) for precise matching
• Parent documents: Larger (e.g., 2000 chars) for complete context
• Map child → parent for retrieval

Available functions:

• create_parent_document_retriever(documents, child_size, parent_size) - Create retriever
• add_documents(documents) - Index documents with parent-child relationship
• retrieve(query, k) - Search children, return parents

Example:

from techniques.en.parent_document import ParentDocumentRetriever

retriever = ParentDocumentRetriever(
    child_chunk_size=400,
    parent_chunk_size=2000
)
retriever.add_documents(documents)

# Searches small chunks, returns full parent context
results = retriever.retrieve("neural network architecture", k=3)

---

27. Multi-Vector Retrieval

Store multiple representations of documents for improved retrieval.

Representation types:

• Original document content
• Generated summaries
• Hypothetical questions the document answers

Available functions:

• create_multi_vector_retriever(documents) - Create retriever with multiple vectors
• generate_summary(document) - Generate document summary
• generate_questions(document) - Generate hypothetical questions
• retrieve(query, k) - Search all representations

Example:

from techniques.en.multi_vector import MultiVectorRetriever

retriever = MultiVectorRetriever()
retriever.add_documents(documents)  # Creates summary + question vectors

# Can match query to summary, questions, or original content
results = retriever.retrieve("How does backpropagation work?", k=3)

---

28. Ensemble Retrieval

Combine multiple retrievers using Reciprocal Rank Fusion with configurable weights.

Components:

• Multiple base retrievers (BM25, Vector, etc.)
• Configurable weights per retriever
• RRF algorithm for score combination

Available functions:

• create_ensemble_retriever(retrievers, weights) - Create ensemble
• reciprocal_rank_fusion(results_list, weights, k) - Combine with RRF

Example:

from techniques.en.ensemble_retrieval import EnsembleRetriever

ensemble = EnsembleRetriever(
    retrievers=[bm25_retriever, vector_retriever, sparse_retriever],
    weights=[0.3, 0.5, 0.2]
)
results = ensemble.retrieve("machine learning optimization", k=5)

---

29. Long Context Strategies

Process documents that exceed typical context windows using various strategies.

Strategies:

• Map-Reduce - Process chunks separately, combine results
• Refine - Iteratively build answer with each chunk
• Map-Rerank - Score each chunk, use best ones
• Stuffing - Fit most relevant content into context

Available functions:

• map_reduce_summarize(chunks) - Summarize with map-reduce
• refine_summarize(chunks) - Iterative refinement
• map_rerank_answer(chunks, question) - Score and select best chunks
• stuffing_with_prioritization(chunks, question, max_context) - Priority-based stuffing

Example:

from techniques.en.long_context import map_reduce_summarize, map_rerank_answer

# Summarize a 50-page document
summary = map_reduce_summarize(document_chunks)

# Answer question using best chunks
answer = map_rerank_answer(
    chunks=document_chunks,
    question="What are the main conclusions?"
)

---

30. Time-Weighted Retrieval

Incorporate temporal relevance with exponential decay to prefer recent documents.

Features:

• Exponential decay function for time weighting
• Configurable decay rate and time units
• Combines semantic similarity with recency

Available functions:

• create_time_weighted_retriever(documents, decay_rate) - Create retriever
• calculate_time_weight(timestamp, decay_rate, time_unit) - Calculate decay weight
• retrieve(query, k, time_weight_factor) - Search with time weighting

Example:

from techniques.en.time_weighted import TimeWeightedRetriever

retriever = TimeWeightedRetriever(
    documents=news_articles,
    decay_rate=0.05,  # Per day
    time_unit="days"
)

# Recent articles score higher
results = retriever.retrieve(
    "AI developments",
    k=5,
    time_weight_factor=0.4  # 40% time, 60% semantic
)

Configuration guide:

Use Case	Decay Rate	Time Unit	Weight Factor
News/Current	0.1-0.5	hours	0.5-0.7
Chat history	0.05-0.1	hours	0.3-0.5
Documentation	0.01-0.05	days	0.2-0.4
Research papers	0.001-0.01	days	0.1-0.3

---

31. MCP Basics

Model Context Protocol (MCP) is an open protocol created by Anthropic to connect AI assistants to external data sources and tools in a standardized way.

Key concepts:

• Resources - Data exposed by the server (files, databases, APIs)
• Tools - Functions that the LLM can invoke
• Prompts - Reusable prompt templates

Example:

from techniques.en.mcp_basics import MCPServerSimulator

server = MCPServerSimulator(name="demo-server")
server.add_tool("search_database", description="Search data", ...)
server.add_resource("file:///config.json", name="Config", ...)

---

32. MCP Server STDIO

STDIO is the most common transport method for local MCP servers. Communication occurs through stdin/stdout.

Use cases:

• Claude Desktop integration
• Command line tools
• Local file access
• Script execution

Example:

# In production, use:
from mcp.server import Server
from mcp.server.stdio import stdio_server

server = Server("my-server")

async def main():
    async with stdio_server() as (read_stream, write_stream):
        await server.run(read_stream, write_stream)

---

33. MCP Server HTTP/SSE

HTTP/SSE is the transport method for remote MCP servers. Enables communication over the network.

Use cases:

• AI-as-a-service APIs
• Enterprise integrations
• Centralized servers
• AI microservices

Example:

# In production, use FastAPI with MCP SDK:
from fastapi import FastAPI
from mcp.server import Server
from mcp.server.sse import SseServerTransport

app = FastAPI()
server = Server("my-http-server")

---

34. Multi-Agent Applications

Multi-agent systems allow multiple AI agents to collaborate to solve complex tasks.

Patterns:

• Pipeline - Sequential processing (Agent1 → Agent2 → Agent3)
• Debate - Agents with different perspectives discuss to reach consensus
• Hierarchical - Agents organized in authority levels (Director → Managers → Workers)
• Orchestrator - Central agent coordinates the others

Frameworks:

• LangGraph (LangChain)
• AutoGen (Microsoft)
• CrewAI
• Swarm (OpenAI)

Example:

from techniques.en.multi_agent import MultiAgentSystem, Agent, AgentRole

system = MultiAgentSystem(name="dev-team")
system.add_agent(Agent(name="Planner", role=AgentRole.PLANNER, ...))
system.add_agent(Agent(name="Coder", role=AgentRole.EXECUTOR, ...))
system.add_agent(Agent(name="Reviewer", role=AgentRole.REVIEWER, ...))

---

35. Prompt Evaluation

Prompt evaluation is essential to ensure quality, consistency, and continuous improvement in LLM applications.

Metrics:

• Relevance - Does response address the question?
• Coherence - Does text have logic and flow?
• Groundedness - Based on facts/context?
• Accuracy - Correct information?
• Safety - Appropriate content?

Tools:

• LangSmith (LangChain)
• LangFuse (Open Source)
• Weights & Biases
• Promptfoo (CLI)
• Phoenix (Arize)

Example:

from techniques.en.prompt_evaluation import PromptEvaluator

evaluator = PromptEvaluator()
result = evaluator.evaluate_relevance(question, answer)
print(f"Relevance: {result.score:.2f}")

---

36. LLM Security

Security measures for LLM applications following OWASP Top 10 guidelines for LLMs.

Security components:

• PromptInjectionDetector - Detects prompt injection attempts
• OutputValidator - Validates and sanitizes LLM outputs
• ContentGuardrail - Enforces content policies
• RateLimiter - Prevents abuse with token bucket algorithm
• SecureLLMWrapper - Combines all security measures

Example:

from techniques.en.llm_security import SecureLLMWrapper

secure_llm = SecureLLMWrapper(
    enable_injection_detection=True,
    enable_output_validation=True,
    rate_limit_rpm=60
)

response = secure_llm.chat("User input here")

---

37. Caching Strategies

Multiple caching approaches to reduce API costs and improve response times.

Cache types:

• ResponseCache - Exact match caching for identical prompts
• EmbeddingCache - Cache embeddings to avoid recomputation
• SemanticCache - Find similar queries using embedding similarity
• ConversationCache - Cache conversation contexts

Example:

from techniques.en.caching_strategies import SemanticCache

cache = SemanticCache(similarity_threshold=0.95)
cache.set("What is Python?", "Python is a programming language...")

# Similar query will return cached response
response = cache.get("Tell me about Python")

---

38. Cost Optimization

Tools and strategies to monitor and reduce LLM API costs.

Components:

• TokenCounter - Count tokens before API calls
• UsageTracker - Track usage and costs over time
• ModelSelector - Choose optimal model based on task complexity

Example:

from techniques.en.cost_optimization import TokenCounter, UsageTracker

counter = TokenCounter()
tokens = counter.count_tokens("Your prompt here")
estimated_cost = counter.estimate_cost(tokens, 500, "gpt-4o-mini")

tracker = UsageTracker()
tracker.daily_budget = 1.00

---

39. AI Testing

Testing frameworks and strategies for non-deterministic LLM outputs.

Testing approaches:

• Property-based validators (contains, length, regex, JSON)
• Semantic similarity testing with embeddings
• LLM-as-Judge evaluation
• Mock clients for CI/CD
• Snapshot testing for regression detection

Example:

from techniques.en.ai_testing import LLMTestRunner, TestCase, ContainsValidator

test = TestCase(
    name="Greeting test",
    prompt="Say hello",
    validators=[ContainsValidator(["hello"])]
)

runner = LLMTestRunner()
result = runner.run_test(test)

---

40. Fine-tuning

Complete workflow for fine-tuning OpenAI models on custom datasets.

Components:

• DatasetGenerator - Create training examples from templates
• DatasetValidator - Validate dataset format and quality
• FineTuningManager - Upload, train, and use fine-tuned models

Example:

from techniques.en.fine_tuning import DatasetGenerator, DatasetValidator

generator = DatasetGenerator(system_prompt="Classify sentiment...")
generator.add_examples_from_pairs([
    ("Great product!", "POSITIVE"),
    ("Terrible experience", "NEGATIVE")
])
generator.export_jsonl("training_data.jsonl")

Token Monitoring

All scripts include automatic token counting to help monitor costs and API usage.

Sample Output

Each LLM call displays the tokens used:

Text: This product is amazing! It exceeded all...
   Tokens - Input: 52 | Output: 3 | Total: 55
Sentiment: POSITIVE

At the end of each script, a total summary is displayed:

============================================================
TOTAL - Zero-Shot Prompting
   Input:  1,234 tokens
   Output: 456 tokens
   Total:  1,690 tokens
============================================================

Project Structure

.
├── .env.example              # Configuration template
├── .gitignore                # Files ignored by Git
├── README.md                 # English documentation
├── README.pt-BR.md           # Portuguese documentation
├── requirements.txt          # Project dependencies
├── config.py                 # Centralized config + Token tracking
├── sample_data/              # Sample data for RAG, Vision, and Context Engineering
│   ├── documents/            # Text documents for RAG and Context Engineering
│   │   ├── ai_handbook.txt
│   │   ├── company_faq.txt
│   │   ├── technical_docs.md
│   │   ├── products_catalog.json   # Product data with metadata (Self-Query)
│   │   ├── news_articles.txt       # Dated articles (Time-Weighted)
│   │   └── long_document.txt       # Large document (Long Context)
│   └── images/               # Images for Vision demos
│       ├── chart.png
│       ├── diagram.png
│       └── photo.jpg
└── techniques/
    ├── en/                   # English examples (45 scripts)
    │   ├── 01_zero_shot.py
    │   ├── ...
    │   ├── 20_meta_prompting.py
    │   ├── 21_advanced_chunking.py
    │   ├── ...
    │   ├── 30_time_weighted.py
    │   ├── 31_mcp_basics.py
    │   ├── ...
    │   ├── 40_fine_tuning.py
    │   ├── 41_agent_skills.py
    │   ├── 42_context_window.py
    │   ├── 43_subagent_orchestration.py
    │   ├── 44_shared_memory.py
    │   └── 45_spec_generation.py
    └── pt-br/                # Portuguese examples (45 scripts)
        ├── 01_zero_shot.py
        ├── ...
        ├── 20_meta_prompting.py
        ├── 21_advanced_chunking.py
        ├── ...
        ├── 30_time_weighted.py
        ├── 31_mcp_basics.py
        ├── ...
        ├── 40_fine_tuning.py
        ├── 41_agent_skills.py
        ├── 42_context_window.py
        ├── 43_subagent_orchestration.py
        ├── 44_shared_memory.py
        └── 45_spec_generation.py

Configuration

The config.py file provides utility functions:

from config import get_llm, get_model_name, TokenUsage

# Create LLM instance with custom temperature
llm = get_llm(temperature=0.7)

# Get configured model name
model = get_model_name()  # e.g., "gpt-4o-mini"

# Create token tracker
tracker = TokenUsage()

# For Ollama (local models)
from config import get_ollama_llm, get_ollama_embeddings, is_ollama_available

if is_ollama_available():
    local_llm = get_ollama_llm(model="llama3.2")
    local_embeddings = get_ollama_embeddings()

# For embeddings
from config import get_embeddings
embeddings = get_embeddings()  # OpenAI embeddings

Understanding Temperature

Temperature is one of the most important parameters when working with LLMs. It controls the randomness and creativity of the model's responses.

What is Temperature?

• Range: 0.0 to 2.0 (most common usage is 0.0 to 1.0)
• Low values (0.0-0.3): More deterministic, focused, and consistent responses
• High values (0.7-1.0+): More creative, diverse, and unpredictable responses

Temperature by Technique

Technique	Temperature	Reason
Zero-Shot Classification	0.0	Consistent results
Chain of Thought	0.0	Accurate reasoning
Few-Shot	0.0 - 0.3	Follow examples
Tree of Thoughts	0.3 - 0.8	Diverse thoughts
Self-Consistency	0.7 - 0.9	Need variation
Self-Refine	0.3 - 0.5	Balanced critique
RAG	0.0 - 0.3	Factual answers
Structured Output	0.0	Consistent schema
Tool Calling	0.0	Reliable tool use
Meta-Prompting	0.5 - 0.7	Creative prompts
Query Transformation	0.3 - 0.7	Creative variations
Contextual Compression	0.0	Accurate extraction
Self-Query (filter gen)	0.0	Precise filters
Multi-Vector (summaries)	0.3	Balanced summaries
Long Context	0.0 - 0.3	Accurate synthesis

Supported Models

OpenAI Models

• gpt-4o - Most capable, most expensive
• gpt-4o-mini - Good cost/performance balance (recommended)
• gpt-4-turbo - Turbo version of GPT-4
• gpt-3.5-turbo - Cheaper, less capable

Ollama Models (Local)

• llama3.2 - Meta's Llama 3 (recommended)
• mistral - Mistral 7B
• codellama - Code-specialized
• phi3 - Microsoft Phi-3

Dependencies

Core Dependencies

• langchain - LLM framework
• langchain-openai - OpenAI integration
• openai - OpenAI API client
• python-dotenv - Environment variables

RAG Dependencies

• chromadb - Vector database
• sentence-transformers - Local embeddings
• pypdf - PDF processing
• unstructured - Document parsing

Context Engineering Dependencies

• rank-bm25 - BM25 keyword search for Hybrid Search

Ollama Dependencies

• langchain-ollama - Ollama integration

Optional Dependencies

• cohere - Cohere reranking
• pillow - Image processing

Usage Tips

Prompt Engineering

1. Start with Zero-Shot - It's the simplest technique and works well for direct tasks.

2. Use CoT for reasoning - Mathematical, logical problems, or those requiring analysis benefit from "think step by step".

3. Few-Shot for specific formats - When you need output in a specific format (JSON, SQL, etc.), provide examples.

4. Self-Consistency for accuracy - When you need high accuracy on reasoning tasks, generate multiple responses and vote.

5. Ollama for privacy - Use local models when data privacy is important or you want to avoid API costs.

6. Structured Output for APIs - When building integrations, use Pydantic models to ensure consistent output.

Context Engineering

7. RAG for knowledge - Use RAG when you need the model to answer based on specific documents.

8. Hybrid Search for precision - Combine BM25 + Vector search when queries contain specific terms or keywords.

9. Advanced Chunking for quality - Choose chunking strategy based on content type (semantic for articles, markdown-aware for docs).

10. Query Transformation for recall - Use HyDE or Multi-Query when initial retrieval quality is low.

11. Contextual Compression for tokens - Compress retrieved documents to reduce token usage while maintaining relevance.

12. Self-Query for structured data - Use when documents have rich metadata that can filter results.

13. Parent-Document for context - When retrieved chunks lack surrounding context, use parent-document retrieval.

14. Long Context for large docs - Use Map-Reduce for summarization, Map-Rerank for Q&A on large documents.

15. Time-Weighted for freshness - Use when document recency matters (news, logs, chat history).

Costs

The scripts make calls to the OpenAI API, which charges per token.

Monitoring Costs

Each script automatically displays:

• Input and output tokens per call
• Total tokens at the end of execution

Approximate prices (January 2025):

Model	Input (1M tokens)	Output (1M tokens)
gpt-4o	$2.50	$10.00
gpt-4o-mini	$0.15	$0.60
gpt-3.5-turbo	$0.50	$1.50

Minimizing Costs

• Use gpt-4o-mini (default) instead of gpt-4o
• Use Ollama for local inference (free)
• Reduce the number of examples in tests
• Monitor token totals displayed at the end of each script

License

MIT