Knowledge Base Injection

# Knowledge Base Injection

Part of [Agent Skills™](https://github.com/itallstartedwithaidea/agent-skills) by [googleadsagent.ai™](https://googleadsagent.ai)

## Description

Knowledge Base Injection is the technique of dynamically injecting domain expertise into an agent's context at the moment it is most relevant, using TF-IDF pattern matching and semantic scoring. Generic language models lack the deep domain knowledge required for specialized tasks like Google Ads optimization, medical coding, or financial compliance. Rather than fine-tuning (expensive, slow, brittle) or bloating system prompts with everything the model might need (wasteful, dilutes attention), Knowledge Base Injection retrieves and injects only the specific patterns relevant to the current task.

This skill is built on the production knowledge base system powering Buddy™ at [googleadsagent.ai™](https://googleadsagent.ai), specifically the `gads-knowledge.js` module containing over 1,000 curated Google Ads optimization patterns. Each pattern includes a trigger condition (when to apply it), a recommendation (what to do), evidence (why it works), and a confidence score. When Buddy™ analyzes a campaign, the knowledge base engine scores all patterns against the current context using TF-IDF and injects the top-K most relevant patterns into the agent's context, transforming a general-purpose model into a domain expert.

The injection system operates on a retrieval-augmented generation (RAG) paradigm, but with a critical distinction: rather than retrieving raw documents, it retrieves structured action patterns with built-in confidence scores and applicability conditions. This produces more actionable, more reliable agent outputs than document-level RAG.

## Use When

- The agent needs domain expertise that general-purpose models lack
- You have a curated knowledge base of patterns, rules, or best practices
- Fine-tuning is too expensive, too slow, or creates model version lock-in
- Different queries require different subsets of domain knowledge
- You want to update the agent's expertise without retraining or redeploying
- The agent must ground its recommendations in verified, authoritative sources

## How It Works

```mermaid
graph TD
    A[Agent Task/Query] --> B[Query Analyzer]
    B --> C[TF-IDF Scorer]
    C --> D[Knowledge Base<br/>1,000+ Patterns]
    D --> E[Ranked Results]
    E --> F{Score > Threshold?}
    F -->|Yes| G[Top-K Selection]
    F -->|No| H[Fallback: Broader Search]
    H --> G
    G --> I[Context Assembly]
    I --> J[Pattern Injection<br/>into Agent Context]
    J --> K[Agent Execution]
    K --> L[Citation Tracking]
    L --> M[Response with Sources]
```

When a task arrives, the query analyzer extracts key terms and concepts. The TF-IDF scorer computes relevance scores between the query and every pattern in the knowledge base. Patterns scoring above the threshold are ranked and the top-K are selected for injection. The selected patterns are assembled into a structured context block with clear formatting and injected into the agent's prompt. After execution, citation tracking links the agent's recommendations back to the specific patterns that informed them, providing auditability.

## Implementation

**Knowledge Pattern Structure:**

```typescript
interface KnowledgePattern {
  id: string;
  category: "bidding" | "targeting" | "creative" | "budget" | "structure" | "general";
  trigger: string;
  recommendation: string;
  evidence: string;
  confidence: number;
  terms: string[];
  tf_idf_vector?: number[];
}

const SAMPLE_PATTERNS: KnowledgePattern[] = [
  {
    id: "bid-001",
    category: "bidding",
    trigger: "Campaign using manual CPC with more than 30 conversions/month",
    recommend