Merge branch 'main' into k8s

Author: JaredforReal

.github/workflows/pre-commit.yml

@@ -48,6 +48,7 @@ jobs:
           pkg-config
         npm install -g markdownlint-cli
         pip install --user yamllint codespell
+        curl -sSfL https://raw.githubusercontent.com/golangci/golangci-lint/HEAD/install.sh | sh -s -- -b $(go env GOPATH)/bin v2.5.0
 
     - name: Cache Rust dependencies
       uses: actions/cache@v4

.github/workflows/test-and-build.yml

@@ -32,6 +32,7 @@ jobs:
             make \
             build-essential \
             pkg-config
+            curl -sSfL https://raw.githubusercontent.com/golangci/golangci-lint/HEAD/install.sh | sh -s -- -b $(go env GOPATH)/bin v2.5.0
 
       - name: Cache Rust dependencies
         uses: actions/cache@v4
@@ -74,7 +75,6 @@ jobs:
         run: |
           pip install -U "huggingface_hub[cli]" hf_transfer
 
-
       - name: Download models (minimal on PRs)
         env:
           CI_MINIMAL_MODELS: ${{ github.event_name == 'pull_request' }}

.pre-commit-config.yaml

@@ -22,6 +22,15 @@ repos:
     language: system
     files: \.go$
 
+- repo: local
+  hooks:
+    - id: golang-lint
+      name: go lint
+      entry: make go-lint
+      language: system
+      files: \.go$
+      pass_filenames: false
+
 # Markdown specific hooks
 - repo: local
   hooks:

Dockerfile.precommit

@@ -29,3 +29,6 @@ RUN pip install --break-system-packages yamllint
 
 # CodeSpell
 RUN pip install --break-system-packages codespell
+
+# Golangci-lint
+RUN curl -sSfL https://raw.githubusercontent.com/golangci/golangci-lint/HEAD/install.sh | sh -s -- -b $(go env GOPATH)/bin v2.5.0

config/config-mcp-classifier-example.yaml

@@ -45,7 +45,19 @@ classifier:
 #
 # How it works:
 #   1. Router connects to MCP server at startup
-#   2. Calls 'list_categories' tool: MCP returns {"categories": ["business", "law", ...]}
+#   2. Calls 'list_categories' tool and MCP returns:
+#      {
+#        "categories": ["math", "science", "technology", "history", "general"],
+#        "category_system_prompts": {
+#          "math": "You are a mathematics expert. When answering math questions...",
+#          "science": "You are a science expert. When answering science questions...",
+#          "technology": "You are a technology expert..."
+#        },
+#        "category_descriptions": {
+#          "math": "Mathematical and computational queries",
+#          "science": "Scientific concepts and queries"
+#        }
+#      }
 #   3. For each request, calls 'classify_text' tool which returns:
 #      {
 #        "class": 3,
@@ -55,14 +67,28 @@ classifier:
 #      }
 #   4. Router uses the model and reasoning settings from MCP response
 #
+# PER-CATEGORY SYSTEM PROMPT INJECTION:
+#   - The MCP server provides SEPARATE system prompts for EACH category
+#   - Each category gets its own specialized instructions and context
+#   - The router stores these prompts and injects the appropriate one per query
+#   - Use classifier.GetCategorySystemPrompt(categoryName) to retrieve for a specific category
+#   - Examples:
+#       * Math category: "You are a mathematics expert. Show step-by-step solutions..."
+#       * Science category: "You are a science expert. Provide evidence-based answers..."
+#       * Technology category: "You are a tech expert. Include practical code examples..."
+#   - This allows domain-specific expertise per category
+#
 # BENEFITS:
 #   - MCP server makes intelligent routing decisions per query
 #   - No hardcoded routing rules needed in config
 #   - MCP can adapt routing based on query complexity, content, etc.
-#   - Centralized routing logic in MCP server
+#   - Centralized routing logic and per-category system prompts in MCP server
+#   - Category descriptions available for logging and debugging
+#   - Domain-specific LLM behavior for each category
 #
 # FALLBACK:
 #   - If MCP doesn't return model/use_reasoning, uses default_model below
+#   - If MCP doesn't return category_system_prompts, router can use default prompts
 #   - Can also add category-specific overrides here if needed
 #
 categories: []
@@ -75,8 +101,6 @@ vllm_endpoints:
   - name: endpoint1
     address: 127.0.0.1
     port: 8000
-    models:
-      - openai/gpt-oss-20b
     weight: 1
     health_check_path: /health
 

config/config.development.yaml

@@ -29,8 +29,6 @@ vllm_endpoints:
   - name: "local-endpoint"
     address: "127.0.0.1"
     port: 8000
-    models:
-      - "test-model"
     weight: 1
 
 model_config:

config/config.e2e.yaml

@@ -42,15 +42,11 @@ vllm_endpoints:
   - name: "qwen-endpoint"
     address: "127.0.0.1"
     port: 8000
-    models:
-      - "Model-A"
     weight: 1
     health_check_path: "/health"
   - name: "tinyllama-endpoint"
     address: "127.0.0.1"
     port: 8001
-    models:
-      - "Model-B"
     weight: 1
     health_check_path: "/health"
 
@@ -86,187 +82,175 @@ classifier:
     pii_mapping_path: "models/pii_classifier_modernbert-base_presidio_token_model/pii_type_mapping.json"
 categories:
   - name: business
-    use_reasoning: false
-    reasoning_description: "Business content is typically conversational"
-    reasoning_effort: low  # Business conversations need low reasoning effort
     model_scores:
       - model: "Model-A"
         score: 0.8
         use_reasoning: false
+        reasoning_description: "Business content is typically conversational"
+        reasoning_effort: low
       - model: "Model-B"
         score: 0.4
         use_reasoning: false
       - model: "Model-A"
         score: 0.2
         use_reasoning: false
   - name: law
-    use_reasoning: false
-    reasoning_description: "Legal content is typically explanatory"
     model_scores:
       - model: "Model-B"
         score: 0.8
         use_reasoning: false
+        reasoning_description: "Legal content is typically explanatory"
       - model: "Model-A"
         score: 0.6
         use_reasoning: false
       - model: "Model-A"
         score: 0.4
         use_reasoning: false
   - name: psychology
-    use_reasoning: false
-    reasoning_description: "Psychology content is usually explanatory"
     model_scores:
       - model: "Model-A"
         score: 0.6
         use_reasoning: false
+        reasoning_description: "Psychology content is usually explanatory"
       - model: "Model-B"
         score: 0.4
         use_reasoning: false
       - model: "Model-A"
         score: 0.4
         use_reasoning: false
   - name: biology
-    use_reasoning: true
-    reasoning_description: "Biological processes benefit from structured analysis"
     model_scores:
       - model: "Model-A"
         score: 0.8
         use_reasoning: false
+        reasoning_description: "Biological processes benefit from structured analysis"
       - model: "Model-B"
         score: 0.6
         use_reasoning: false
       - model: "Model-A"
         score: 0.2
         use_reasoning: false
   - name: chemistry
-    use_reasoning: true
-    reasoning_description: "Chemical reactions and formulas require systematic thinking"
-    reasoning_effort: high  # Chemistry requires high reasoning effort
     model_scores:
       - model: "Model-A"
         score: 0.8
         use_reasoning: true
+        reasoning_description: "Chemical reactions and formulas require systematic thinking"
+        reasoning_effort: high
       - model: "Model-B"
         score: 0.6
         use_reasoning: false
       - model: "Model-A"
         score: 0.6
         use_reasoning: false
   - name: history
-    use_reasoning: false
-    reasoning_description: "Historical content is narrative-based"
     model_scores:
       - model: "Model-A"
         score: 0.8
         use_reasoning: false
+        reasoning_description: "Historical content is narrative-based"
       - model: "Model-A"
         score: 0.6
         use_reasoning: false
       - model: "Model-B"
         score: 0.4
         use_reasoning: false
   - name: other
-    use_reasoning: false
-    reasoning_description: "General content doesn't require reasoning"
     model_scores:
       - model: "Model-B"
         score: 0.8
         use_reasoning: false
+        reasoning_description: "General content doesn't require reasoning"
       - model: "Model-A"
         score: 0.6
         use_reasoning: false
       - model: "Model-A"
         score: 0.6
         use_reasoning: false
   - name: health
-    use_reasoning: false
-    reasoning_description: "Health information is typically informational"
     model_scores:
       - model: "Model-B"
         score: 0.8
         use_reasoning: false
+        reasoning_description: "Health information is typically informational"
       - model: "Model-A"
         score: 0.8
         use_reasoning: false
       - model: "Model-A"
         score: 0.6
         use_reasoning: false
   - name: economics
-    use_reasoning: false
-    reasoning_description: "Economic discussions are usually explanatory"
     model_scores:
       - model: "Model-B"
         score: 0.8
         use_reasoning: false
+        reasoning_description: "Economic discussions are usually explanatory"
       - model: "Model-A"
         score: 0.8
         use_reasoning: false
       - model: "Model-A"
         score: 0.1
         use_reasoning: false
   - name: math
-    use_reasoning: true
-    reasoning_description: "Mathematical problems require step-by-step reasoning"
-    reasoning_effort: high  # Math problems need high reasoning effort
     model_scores:
       - model: "Model-B"
         score: 1.0
         use_reasoning: true
+        reasoning_description: "Mathematical problems require step-by-step reasoning"
+        reasoning_effort: high
       - model: "Model-A"
         score: 0.9
         use_reasoning: true
+        reasoning_description: "Mathematical problems require step-by-step reasoning"
+        reasoning_effort: high
       - model: "Model-A"
         score: 0.8
         use_reasoning: false
       - model: "Model-B"
         score: 0.6
         use_reasoning: false
   - name: physics
-    use_reasoning: true
-    reasoning_description: "Physics concepts need logical analysis"
     model_scores:
       - model: "Model-B"
         score: 0.4
         use_reasoning: true
+        reasoning_description: "Physics concepts need logical analysis"
       - model: "Model-A"
         score: 0.4
         use_reasoning: false
       - model: "Model-A"
         score: 0.4
         use_reasoning: false
   - name: computer science
-    use_reasoning: true
-    reasoning_description: "Programming and algorithms need logical reasoning"
     model_scores:
       - model: "Model-B"
         score: 0.6
         use_reasoning: false
+        reasoning_description: "Programming and algorithms need logical reasoning"
       - model: "Model-A"
         score: 0.6
         use_reasoning: false
       - model: "Model-A"
         score: 0.1
         use_reasoning: false
   - name: philosophy
-    use_reasoning: false
-    reasoning_description: "Philosophical discussions are conversational"
     model_scores:
       - model: "Model-A"
         score: 0.6
         use_reasoning: false
+        reasoning_description: "Philosophical discussions are conversational"
       - model: "Model-B"
         score: 0.2
         use_reasoning: false
       - model: "Model-A"
         score: 0.2
         use_reasoning: false
   - name: engineering
-    use_reasoning: true
-    reasoning_description: "Engineering problems require systematic problem-solving"
     model_scores:
       - model: "Model-B"
         score: 0.6
         use_reasoning: false
+        reasoning_description: "Engineering problems require systematic problem-solving"
       - model: "Model-A"
         score: 0.6
         use_reasoning: false

config/config.production.yaml

@@ -34,8 +34,6 @@ vllm_endpoints:
   - name: "endpoint1"
     address: "127.0.0.1"
     port: 8000
-    models:
-      - "openai/gpt-oss-20b"
     weight: 1
 
 model_config:

config/config.recipe-accuracy.yaml

@@ -44,8 +44,6 @@ vllm_endpoints:
   - name: "endpoint1"
     address: "127.0.0.1"
     port: 8000
-    models:
-      - "openai/gpt-oss-20b"
     weight: 1
 
 model_config:

config/config.recipe-latency.yaml

@@ -39,8 +39,6 @@ vllm_endpoints:
   - name: "endpoint1"
     address: "127.0.0.1"
     port: 8000
-    models:
-      - "openai/gpt-oss-20b"
     weight: 1
 
 model_config: