07. Context Engineering

Developer A: "Is this query efficient?"
→ "It depends on your schema, indexes, and data size..."

Developer B: "Is this query efficient?
   Context: MySQL 8.0, 10 million rows in 'orders' table,
   index on (customer_id, order_date).
   Query: SELECT * FROM orders WHERE customer_id = 5"
→ "Yes, the index makes this efficient. However, SELECT *
   reads all columns. Consider selecting only needed columns."

flowchart LR
    subgraph NOCONTEXT["Without Context"]
        W["Question"] --> W1["Generic Answer"]
    end

    subgraph CONTEXT["With Context"]
        C["Question + Context"] --> C1["Search relevant knowledge"]
        C1 --> C2["Apply to specific situation"]
        C2 --> C3["Specific, actionable Answer"]
    end

    style NOCONTEXT fill:#ef4444,color:#fff
    style CONTEXT fill:#22c55e,color:#fff

flowchart TD
    subgraph GOOD["Good Context"]
        G1["Relevant\nOnly what's needed"] --> G2["Specific\nExact details"]
        G2 --> G3["Structured\nOrganized clearly"]
        G3 --> G4["Recent\nPlaced near instruction"]
    end

    subgraph POOR["Poor Context"]
        P1["Irrelevant\nIncludes unrelated info"] --> P2["Vague\nMissing specifics"]
        P2 --> P3["Messy\nUnorganized dump"]
        P3 --> P4["Buried\nLost in the prompt"]
    end

    style GOOD fill:#22c55e,color:#fff
    style POOR fill:#ef4444,color:#fff

flowchart LR
    subgraph WINDOW["Context Window"]
        INSTRUCTION["Instruction\n(~100 tokens)"] --> CONTEXT["Context\n(~variable tokens)"]
        CONTEXT --> DATA["User Input\n(~variable)"]
        DATA --> OUTPUT["Model Output\n(~variable)"]
    end

    style WINDOW fill:#3b82f6,color:#fff
    style INSTRUCTION fill:#f59e0b,color:#fff
    style CONTEXT fill:#22c55e,color:#fff
    style DATA fill:#8b5cf6,color:#fff
    style OUTPUT fill:#ec4899,color:#fff

Total Context = System Prompt + Conversation History + User Input + Context Documents + Expected Output

Example:
System Prompt: 500 tokens
Conversation: 2,000 tokens (4 turns)
User Input: 200 tokens
Context Documents: 10,000 tokens
Expected Output: 500 tokens
──────────────────────────────────────
Total: 13,200 tokens (well within 128K limit)

flowchart TD
    subgraph PLACEMENT["Context Placement Strategy"]
        CRITICAL["Critical Instructions\nPlace near beginning AND near end"] --> IMPORTANT["Important Context\nPlace near beginning"]
        IMPORTANT --> REFERENCE["Reference Material\nPlace in middle"]
        REFERENCE --> RECENT["Recent/Updated Info\nPlace near end (recency bias)"]
    end

    style CRITICAL fill:#ef4444,color:#fff
    style IMPORTANT fill:#f59e0b,color:#fff
    style REFERENCE fill:#3b82f6,color:#fff
    style RECENT fill:#22c55e,color:#fff

High Retention ────┐                  ┌──── High Retention
                   │                  │
                   ▼                  ▼
              [Beginning] [Middle] [End]
                   │                  ▲
                   └──────────────────┘
                    Low retention here

flowchart TD
    COMPRESSION["Context Compression"] --> S1["Summarization\nCondense documents to key points"]
    COMPRESSION --> S2["Chunking\nSplit into smaller pieces"]
    COMPRESSION --> S3["Filtering\nRemove irrelevant content"]
    COMPRESSION --> S4["Extraction\nExtract only relevant sections"]
    COMPRESSION --> S5["Ranking\nPrioritize by relevance"]

    style COMPRESSION fill:#8b5cf6,color:#fff

flowchart LR
    subgraph TURNS["Conversation Turns"]
        T1["Turn 1"] --> T2["Turn 2"] --> T3["..."] --> T4["Turn 50"] --> T5["Turn 51"]
    end

    subgraph WINDOW["Sliding Window (last 10 turns)"]
        T4 --> W1["Keep"]
        T5 --> W2["Keep"]
        SUMMMARY["Summary of Turns 1-40"] --> W3["Summary"]
    end

    style WINDOW fill:#22c55e,color:#fff

CONTEXT:
── PROJECT ──
Name: [project name]
Stack: [tech stack]
Stage: [development stage]

── CURRENT TASK ──
Branch: [git branch]
Files Changed: [list of files]
PR Description: [summary]

── RELEVANT CODE ──
[code snippets]

── CONSTRAINTS ──
[deadlines, requirements, limitations]

Turn 1: "You are a senior engineer. Here's the full context..."
Turn 10: "Remember, you're a senior engineer. The project uses TypeScript."
Turn 20: "Quick reminder: this is a production system. Performance is critical."

❌ Without Context:
"Review this pull request."
→ Generic feedback, might miss project-specific concerns

✅ With Context:
"Review this pull request.
Context:
- Project: Real-time chat application
- Stack: Node.js, Socket.io, Redis, TypeScript
- PR Size: 3 files changed, 120 additions
- Concern: We've been having memory leak issues in production
- Team Convention: Every function must have unit tests

[PR Code Here]"
→ Focused review that considers project context and known issues

❌ Without Context:
"My app is slow. What should I do?"
→ 20 generic optimization suggestions

✅ With Context:
"My app is slow.
Context:
- Hosted on: AWS t2.micro (1 vCPU, 1GB RAM)
- Stack: Node.js, Express, MongoDB
- Issue: Response time increased from 200ms to 2s after deploying last update
- Last change: Added image processing endpoint
- Traffic: ~100 req/s during peak

What should I do?"
→ Specific diagnosis: the image processing is likely overwhelming your small instance