update

Signed-off-by: bitliu <[email protected]>

Author: Xunzhuo

_posts/2025-11-18-signal-decision.md

@@ -78,22 +78,7 @@ The Signal-Decision Architecture fundamentally reimagines semantic routing by se
 
 ### Architecture Overview
 
-```mermaid
-graph TB
-    A[User Request] --> B[Multi-Signal Extraction]
-    B --> C[Keyword Signals]
-    B --> D[Embedding Signals]
-    B --> E[Domain Signals]
-    C --> F[Decision Engine]
-    D --> F
-    E --> F
-    F --> G{Decision Matched?}
-    G -->|Yes| H[Plugin Chain]
-    G -->|No| I[Default Model]
-    H --> J[Model Selection]
-    I --> K[Response]
-    J --> K
-```
+![](/assets/figures/semantic-router/signal-1.png)
 
 The new architecture introduces three key innovations:
 
@@ -103,65 +88,15 @@ The new architecture introduces three key innovations:
 
 ### Complete Request Flow
 
-```mermaid
-sequenceDiagram
-    participant User
-    participant Gateway as AI Gateway
-    participant Extractor as Signal Extractor
-    participant Engine as Decision Engine
-    participant Plugins as Plugin Chain
-    participant Model as Selected Model
-
-    User->>Gateway: Send Query
-    Gateway->>Extractor: Extract Signals
-
-    par Parallel Signal Extraction
-        Extractor->>Extractor: Keyword Matching
-        Extractor->>Extractor: Embedding Similarity
-        Extractor->>Extractor: Domain Classification
-    end
-
-    Extractor->>Engine: Signals: {keyword, embedding, domain}
-    Engine->>Engine: Evaluate All Decisions
-    Engine->>Engine: Select by Priority
-    Engine->>Plugins: Execute Plugin Chain
-
-    Plugins->>Plugins: 1. Jailbreak Detection
-    Plugins->>Plugins: 2. PII Protection
-    Plugins->>Plugins: 3. Semantic Cache Check
-
-    alt Cache Hit
-        Plugins->>User: Return Cached Response
-    else Cache Miss
-        Plugins->>Plugins: 4. System Prompt Injection
-        Plugins->>Plugins: 5. Header Mutation
-        Plugins->>Model: Route Request
-        Model->>Plugins: Model Response
-        Plugins->>Plugins: Cache Response
-        Plugins->>User: Return Response
-    end
-```
+![](/assets/figures/semantic-router/signal-2.png)
 
 ## Core Concepts
 
 ### Signals: Multi-Dimensional Prompt Analysis
 
 Instead of relying solely on domain classification, the Signal-Decision Architecture extracts three complementary types of signals from each user query. Each signal type leverages different AI/ML techniques and serves distinct purposes in the routing decision process.
 
-```mermaid
-graph TB
-    A[User Query] --> B[Signal Extraction Layer]
-    B --> C[Keyword Extractor<br/>Regex Matching]
-    B --> D[Embedding Model<br/>Sentence Transformers]
-    B --> E[Domain Classifier<br/>MMLU + LoRA]
-    C --> F[Keyword Signals<br/>urgency, security, etc.]
-    D --> G[Embedding Signals<br/>intent similarity scores]
-    E --> H[Domain Signals<br/>computer_science, etc.]
-    F --> I[Decision Engine]
-    G --> I
-    H --> I
-    I --> J[Selected Decision]
-```
+![](/assets/figures/semantic-router/signal-3.png)
 
 #### Keyword Signals: Interpretable Pattern Matching
 
@@ -222,22 +157,7 @@ Domain signals use MMLU-trained classification models to identify the academic o
 
 This enables organizations to extend domain classification to their specific verticals while maintaining the base model's general knowledge.
 
-```mermaid
-graph TB
-    A[User Query:<br/>'Review this medical imaging protocol'] --> B[Domain Classifier]
-    B --> C{Base MMLU Model}
-    C --> D[Detect: Healthcare Domain]
-    D --> E{Load LoRA Adapter}
-    E --> F[medical_imaging LoRA]
-    E --> G[clinical_trials LoRA]
-    E --> H[pharmaceutical_research LoRA]
-    F --> I[Fine-grained Classification:<br/>medical_imaging]
-    I --> J[Route to Specialized Model:<br/>medical-imaging-expert]
-
-    style F fill:#e1f5ff
-    style I fill:#c3e6cb
-    style J fill:#d4edda
-```
+![](/assets/figures/semantic-router/signal-4.png)
 
 **Use Cases**:
 
@@ -297,34 +217,7 @@ Each decision consists of:
 
 #### Decision Evaluation Flow
 
-```mermaid
-graph TB
-    A[Extracted Signals] --> B{Evaluate All Decisions}
-    B --> C[Decision 1: Priority 100<br/>Rule: urgency AND security AND cs]
-    B --> D[Decision 2: Priority 80<br/>Rule: code-review AND cs]
-    B --> E[Decision 3: Priority 60<br/>Rule: architecture-design OR cs]
-    C --> F{Match?}
-    D --> G{Match?}
-    E --> H{Match?}
-    F -->|Yes| I[Matched: Priority 100]
-    F -->|No| J[Not Matched]
-    G -->|Yes| K[Matched: Priority 80]
-    G -->|No| L[Not Matched]
-    H -->|Yes| M[Matched: Priority 60]
-    H -->|No| N[Not Matched]
-    I --> O{Multiple Matches?}
-    K --> O
-    M --> O
-    J --> P{Any Match?}
-    L --> P
-    N --> P
-    O -->|Yes| Q[Select Highest Priority]
-    O -->|No| R[Use Single Match]
-    P -->|No| S[Fallback to Default Model]
-    Q --> T[Execute Plugin Chain]
-    R --> T
-    S --> U[Route to Default Model]
-```
+![](/assets/figures/semantic-router/signal-5.png)
 
 When multiple decisions match, the system selects the one with the highest priority. If no decisions match, the system falls back to the default model.
 
@@ -344,20 +237,7 @@ Plugins execute in the configured order, with each plugin able to modify the req
 
 #### Plugin Chain Execution Flow
 
-```mermaid
-graph LR
-    A[Incoming Request] --> B[Plugin 1: Jailbreak Detection]
-    B -->|Pass| C[Plugin 2: PII Protection]
-    B -->|Block| Z[Return Error]
-    C -->|Redact PII| D[Plugin 3: Semantic Cache]
-    D -->|Cache Hit| Y[Return Cached Response]
-    D -->|Cache Miss| E[Plugin 4: System Prompt Injection]
-    E -->|Add Prompt| F[Plugin 5: Header Mutation]
-    F -->|Add Headers| G[Route to Model]
-    G --> H[Model Response]
-    H --> I[Cache Response]
-    I --> J[Return to User]
-```
+![](/assets/figures/semantic-router/signal-6.png)
 
 ## Scaling from 14 to Unlimited
 
@@ -636,25 +516,7 @@ This configuration demonstrates:
 
 ### Dynamic Configuration Flow
 
-```mermaid
-sequenceDiagram
-    participant User
-    participant K8s
-    participant Controller
-    participant Router
-    participant Model
-
-    User->>K8s: Apply IntelligentRoute CRD
-    K8s->>Controller: Watch CRD changes
-    Controller->>Controller: Convert CRD to internal config
-    Controller->>Router: Update routing rules
-    Router->>Router: Reload configuration
-    Note over Router: No downtime
-    User->>Router: Send request
-    Router->>Router: Evaluate with new rules
-    Router->>Model: Route to selected model
-    Model->>User: Return response
-```
+![](/assets/figures/semantic-router/signal-7.png)
 
 The Kubernetes-native design enables: