Qwen + RAG¶

RAG (Retrieval-Augmented Generation) combines your documents with the LLM to provide accurate, grounded responses.

RAG Architecture¶

flowchart TD
    A[User Query] --> B[Query Embedding]
    B --> C[Vector Search]
    C --> D[Knowledge Base]
    D --> E[Relevant Documents]
    E --> F[Context Assembly]
    F --> G[LLM Generation]
    G --> H[Response + Citations]

    style A fill:#e3f2fd
    style D fill:#fff3e0
    style H fill:#c8e6c9

Document Processing Flow¶

flowchart LR
    A[Documents] --> B[Chunking]
    B --> C[Embedding]
    C --> D[Vector Store]
    D --> E[Ready for Query]

    style D fill:#c8e6c9

Query Processing Pipeline¶

sequenceDiagram
    participant U as User
    participant Q as Query Engine
    participant V as Vector DB
    participant L as LLM

    U->>Q: "What is the SLA for P1?"
    Q->>V: Embed query
    V-->>Q: Top-K similar chunks
    Q->>Q: Assemble context
    Q->>L: Prompt + Context
    L-->>Q: Grounded response
    Q-->>U: "P1 SLA is 4 hours..."

RAG vs Direct LLM¶

Aspect	Direct LLM	RAG
Knowledge	Training data only	Your documents
Accuracy	May hallucinate	Grounded in facts
Updates	Retrain needed	Update knowledge base
Citations	Not available	Source citations

Key Components¶

graph TD
    A[RAG System] --> B[Document Loader]
    A --> C[Text Splitter]
    A --> D[Embeddings]
    A --> E[Vector Store]
    A --> F[Retriever]
    A --> G[Generator LLM]

    B --> C
    C --> D
    D --> E
    E --> F
    F --> G

    style A fill:#e8eaf6
    style G fill:#c8e6c9

Use Cases¶

Policy Compliance: Answer questions from company policies
Technical Support: Grounded in troubleshooting guides
HR Queries: Based on employee handbook
Training: Onboarding documentation

Quick Start¶

# 1. Prepare your documents
# 2. Index them into vector store
python qwen-rag/index_documents.py

# 3. Query the knowledge base
python qwen-rag/query_knowledge.py

Performance Tips¶

Factor	Recommendation
Chunk size	500-1000 tokens
Top-K results	3-5 documents
Embedding model	Use same for index/query

ISP Sales Bot (isp_sales.py)¶

This is an intelligent sales assistant that generates personalized Bangla/English sales pitches for ISP customers. It uses keyword matching and LLM generation.

Features¶

Customer Profiles: Track customer's current plan and requirements
Product Catalog: Multiple ISP packages with keywords
Smart Matching: Keyword-based package recommendation
Bilingual Output: Bangla + English sales pitches

Architecture¶

graph TB
    subgraph "Customer Layer"
        A[Customer Select] --> B[Get Profile]
    end

    subgraph "Matching Layer"
        B --> C[Extract Needs]
        C --> D[Keyword Matching]
        D --> E[Best Package]
    end

    subgraph "Generation Layer"
        E --> F[Build Prompt]
        F --> G[LLM - Qwen 2.5]
        G --> H[Sales Pitch]
    end

    subgraph "UI Layer"
        H --> I[Display Result]
        I --> J[Show Logic]
    end

Workflow¶

sequenceDiagram
    participant User as User
    participant App as Streamlit App
    participant LLM as Qwen 2.5 (LM Studio)

    User->>App: Select Customer
    User->>App: Click "Generate Sales Pitch"
    App->>App: Find Best Package (Keyword Matching)
    App->>LLM: Send Prompt with Context
    LLM-->>App: Receive Sales Pitch Response
    App->>User: Show Proposal

Sample Customer Profiles¶

Customer	Current Plan	Needs	Best Match
Arif Ahmed	5 Mbps	YouTube, browsing	Standard Plan (P1)
Sultana Razia	10 Mbps	Netflix, 4K movies	Entertainment Pro (P2)
Tanvir Hasan	20 Mbps	Work from home, VPN	Business Executive (P3)
Farhana Islam	None (New)	Social media, research	Student Starter (P4)

Product Catalog¶

Package	Speed	Keywords
P1: Standard	10 Mbps	YouTube, browsing
P2: Entertainment Pro	20 Mbps	Netflix, 4K, streaming
P3: Business Executive	100 Mbps	VPN, business, Zoom
P4: Student Starter	5 Mbps	student, affordable

Running the App¶

cd c:\Downloads\classifier-app\qwen-rag
streamlit run isp_sales.py

Access at: http://localhost:8501

Requirements¶

LM Studio running at http://localhost:1234
Qwen 2.5 1.5B model loaded
Streamlit installed

Benefits¶

Benefit	Description
Personalization	Tailored pitches based on customer needs
Bilingual	Bangla builds trust, English explains details
Speed	Real-time generation with local LLM
Privacy	All processing happens locally