Skip to content

Core Concepts

Understanding the key concepts behind Prompt Amplifier.

What is Prompt Amplification?

Prompt Amplification is the process of transforming short, ambiguous user inputs into detailed, structured prompts that LLMs can execute effectively.

The Problem

Users often provide vague prompts:

  • ❌ "How's the project going?"
  • ❌ "Summarize the data"
  • ❌ "Help me with this"

LLMs struggle to provide useful responses without:

  • Context about what "project" or "data" refers to
  • Understanding of what metrics matter
  • Knowledge of the expected output format

The Solution

Prompt Amplifier uses your knowledge base to:

  1. Understand the user's intent
  2. Retrieve relevant context
  3. Generate a comprehensive prompt
User Input: "How's the deal going?"
                    ↓
        [Knowledge Base Search]
                    ↓
          [Context Retrieval]
                    ↓
         [Prompt Generation]
                    ↓
Expanded Prompt: "Generate a Deal Health Assessment report with 
                 the following structure:
                 1. Executive Summary...
                 2. Key Metrics Table...
                 3. Risk Factors...
                 ..."

Architecture

graph TB
    subgraph "Document Ingestion"
        A[Documents] --> B[Loader]
        B --> C[Chunker]
        C --> D[Chunks]
    end

    subgraph "Embedding & Storage"
        D --> E[Embedder]
        E --> F[Vectors]
        F --> G[Vector Store]
    end

    subgraph "Query Processing"
        H[User Query] --> I[Embedder]
        I --> J[Query Vector]
        J --> K[Retriever]
        G --> K
        K --> L[Relevant Chunks]
    end

    subgraph "Prompt Generation"
        H --> M[LLM Generator]
        L --> M
        M --> N[Expanded Prompt]
    end

Key Components

1. Documents & Loaders

Documents are your knowledge base - files containing information that provides context.

Loaders read different file formats:

Loader Formats Use Case
TxtLoader .txt Plain text
PDFLoader .pdf Documents, reports
DocxLoader .docx Word documents
CSVLoader .csv Structured data
JSONLoader .json API responses
ExcelLoader .xlsx Spreadsheets

2. Chunkers

Chunkers split documents into smaller pieces for efficient retrieval.

Document: "This is a long document with many paragraphs..."
              ↓ Chunker
Chunks: ["This is a long...", "...document with...", "...many paragraphs"]

Why chunk?

  • Embeddings have token limits
  • Smaller chunks = more precise retrieval
  • Overlapping chunks preserve context

3. Embedders

Embedders convert text into numerical vectors (embeddings).

Text: "Customer satisfaction increased"
              ↓ Embedder
Vector: [0.12, -0.45, 0.78, ..., 0.23]  # 384 dimensions

Types:

Type Examples Pros Cons
Sparse TF-IDF, BM25 Fast, exact matches No semantic understanding
Dense Sentence Transformers, OpenAI Semantic understanding Slower, may need API

4. Vector Stores

Vector Stores persist embeddings for efficient similarity search.

# Add embeddings
store.add(chunks, embeddings)

# Search for similar
results = store.search(query_embedding, k=5)

Options:

Store Type Persistence Best For
MemoryStore In-memory No Testing
ChromaStore Local DB Yes Development
FAISSStore Local index Yes Large datasets

5. Retrievers

Retrievers find relevant chunks for a given query.

query = "deal health"
results = retriever.retrieve(query, k=5)
# Returns top 5 most relevant chunks

Strategies:

Strategy How it Works Best For
Vector Cosine similarity General queries
Hybrid BM25 + Vector Keyword + semantic
MMR Diversity-optimized Varied results

6. Generators

Generators use LLMs to create expanded prompts from context.

generator.generate(
    query="deal health",
    context=[chunk1, chunk2, chunk3]
)
# Returns detailed prompt

Providers:

Provider Models Quality Cost
OpenAI GPT-4o, GPT-4o-mini Excellent $$
Anthropic Claude 3.5, Claude 3 Excellent $$
Google Gemini 2.0, Gemini 1.5 Very Good $

Data Flow

1. Ingestion Flow

PDF/DOCX/CSV → Loader → Document → Chunker → Chunks → Embedder → Vectors → Store

2. Query Flow

User Query → Embedder → Query Vector → Retriever → Top-K Chunks → Generator → Expanded Prompt

Key Metrics

Expansion Ratio

How much the prompt was expanded:

ratio = len(expanded_prompt) / len(original_prompt)
# e.g., 50x means the output is 50 times longer

Retrieval Score

Similarity between query and retrieved chunks (0-1):

  • > 0.7 = Highly relevant
  • 0.5 - 0.7 = Somewhat relevant
  • < 0.5 = Low relevance

Generation Time

Time to generate the expanded prompt:

  • OpenAI: ~2-5 seconds
  • Anthropic: ~1-3 seconds
  • Google: ~2-4 seconds

Best Practices

1. Quality Knowledge Base

  • Use relevant, up-to-date documents
  • Include examples of good prompts
  • Cover different scenarios

2. Appropriate Chunk Size

  • Too small: loses context
  • Too large: retrieves irrelevant info
  • Sweet spot: 256-512 characters

3. Choose the Right Embedder

  • TF-IDF: Fast, good for exact keywords
  • Sentence Transformers: Best free option
  • OpenAI: Highest quality, but costs money

4. Use Persistent Storage

For production, always use ChromaDB or FAISS:

vectorstore=ChromaStore(
    collection_name="prod_docs",
    persist_directory="/data/embeddings"
)

Next Steps