vllm-sr 0.1.0b2.dev20260126190945__tar.gz → 0.1.0b2.dev20260202073049__tar.gz

{vllm_sr-0.1.0b2.dev20260126190945/vllm_sr.egg-info → vllm_sr-0.1.0b2.dev20260202073049}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: vllm-sr
-Version: 0.1.0b2.dev20260126190945
+Version: 0.1.0b2.dev20260202073049
 Summary: vLLM Semantic Router - Intelligent routing for Mixture-of-Models
 Author: vLLM-SR Team
 License: Apache-2.0

{vllm_sr-0.1.0b2.dev20260126190945 → vllm_sr-0.1.0b2.dev20260202073049}/cli/merger.py

@@ -185,6 +185,37 @@ def translate_context_signals(context_rules: list) -> list:
     return rules
 
 
+def translate_complexity_signals(complexity_rules: list) -> list:
+    """
+    Translate complexity signals to router format.
+
+    Args:
+        complexity_rules: List of ComplexityRule objects
+
+    Returns:
+        list: Router complexity rules
+    """
+    rules = []
+    for signal in complexity_rules:
+        rule = {
+            "name": signal.name,
+            "threshold": signal.threshold,
+            "hard": {"candidates": signal.hard.candidates},
+            "easy": {"candidates": signal.easy.candidates},
+        }
+        if signal.description:
+            rule["description"] = signal.description
+        if signal.composer:
+            rule["composer"] = {
+                "operator": signal.composer.operator,
+                "conditions": [
+                    {"type": c.type, "name": c.name} for c in signal.composer.conditions
+                ],
+            }
+        rules.append(rule)
+    return rules
+
+
 def translate_external_models(external_models: list) -> list:
     """
     Translate external models to router format.
@@ -441,6 +472,14 @@ def merge_configs(user_config: UserConfig, defaults: Dict[str, Any]) -> Dict[str
             )
             log.info(f"  Added {len(user_config.signals.context)} context signals")
 
+        if user_config.signals.complexity and len(user_config.signals.complexity) > 0:
+            merged["complexity_rules"] = translate_complexity_signals(
+                user_config.signals.complexity
+            )
+            log.info(
+                f"  Added {len(user_config.signals.complexity)} complexity signals"
+            )
+
         # Translate domains to categories
         if user_config.signals.domains:
             merged["categories"] = translate_domains_to_categories(

{vllm_sr-0.1.0b2.dev20260126190945 → vllm_sr-0.1.0b2.dev20260202073049}/cli/models.py

@@ -85,6 +85,28 @@ class ContextRule(BaseModel):
     description: Optional[str] = None
 
 
+class ComplexityCandidates(BaseModel):
+    """Complexity candidates configuration."""
+
+    candidates: List[str]
+
+
+class ComplexityRule(BaseModel):
+    """Complexity-based signal configuration using embedding similarity.
+
+    The composer field allows filtering based on other signals (e.g., only apply
+    code_complexity when domain is "computer_science"). This is evaluated after
+    all signals are computed in parallel, enabling signal dependencies.
+    """
+
+    name: str
+    threshold: float = 0.1
+    hard: ComplexityCandidates
+    easy: ComplexityCandidates
+    description: Optional[str] = None
+    composer: Optional["Rules"] = None  # Forward reference, defined below
+
+
 class Signals(BaseModel):
     """All signal configurations."""
 
@@ -97,6 +119,7 @@ class Signals(BaseModel):
     language: Optional[List[Language]] = []
     latency: Optional[List[Latency]] = []
     context: Optional[List[ContextRule]] = []
+    complexity: Optional[List[ComplexityRule]] = []
 
 
 class Condition(BaseModel):

{vllm_sr-0.1.0b2.dev20260126190945 → vllm_sr-0.1.0b2.dev20260202073049}/cli/templates/config.template.yaml

@@ -186,8 +186,8 @@ signals:
       max_tpot: 0.15  # 150ms per token
       description: "For standard applications with moderate latency tolerance"
 
-  # context_rules - Context length signals (Token Count)
-  context_rules:
+  # context - Context length signals (Token Count)
+  context:
     - name: "low_token_count"
       min_tokens: "0"
       max_tokens: "1K"
@@ -197,6 +197,60 @@ signals:
       max_tokens: "128K"
       description: "Long requests requiring large context window"
 
+  # complexity - Complexity signals (Embedding-based difficulty detection)
+  # IMPORTANT: It is strongly recommended to configure a composer for each complexity rule
+  # to filter based on other signals (e.g., domain). This prevents misclassification where
+  # a math question might match code_complexity or vice versa.
+  complexity:
+    - name: "code_complexity"
+      composer:
+        operator: "AND"
+        conditions:
+          - type: "domain"
+            name: "computer science"
+      threshold: 0.1
+      hard:
+        candidates:
+          - "design distributed system"
+          - "implement consensus algorithm"
+          - "optimize for scale"
+          - "architect microservices"
+          - "fix race condition"
+          - "implement garbage collector"
+      easy:
+        candidates:
+          - "print hello world"
+          - "loop through array"
+          - "read file"
+          - "sort list"
+          - "string concatenation"
+          - "simple function"
+      description: "Detects code complexity level"
+    - name: "math_complexity"
+      composer:
+        operator: "AND"
+        conditions:
+          - type: "domain"
+            name: "math"
+      threshold: 0.1
+      hard:
+        candidates:
+          - "prove mathematically"
+          - "derive the equation"
+          - "formal proof"
+          - "solve differential equation"
+          - "prove by induction"
+          - "analyze convergence"
+      easy:
+        candidates:
+          - "what is 2+2"
+          - "simple arithmetic"
+          - "basic calculation"
+          - "count numbers"
+          - "add two numbers"
+          - "multiply values"
+      description: "Detects mathematical complexity level"
+
 # Decisions - Routing logic
 decisions:
   # Highest priority: Preference-based routing via external LLM

{vllm_sr-0.1.0b2.dev20260126190945 → vllm_sr-0.1.0b2.dev20260202073049}/cli/templates/router-defaults.yaml

@@ -10,6 +10,7 @@
 #   "models/mom-feedback-detector": "llm-semantic-router/feedback-detector"
 #   "models/mom-embedding-pro": "Qwen/Qwen3-Embedding-0.6B"
 #   "models/mom-embedding-flash": "google/embeddinggemma-300m"
+#   "models/mom-embedding-ultra": "llm-semantic-router/mmbert-embed-32k-2d-matryoshka"
 
 # Response API Configuration
 # Enables OpenAI Response API support with conversation chaining
@@ -19,6 +20,14 @@ response_api:
   ttl_seconds: 86400       # 24 hours
   max_responses: 1000
 
+# Router Replay Configuration (System-Level)
+# Provides storage backend configuration for router_replay plugin
+# Per-decision settings (max_records, capture settings) are configured via router_replay plugin
+router_replay:
+  store_backend: "memory"  # Options: "memory", "redis", "postgres", "milvus"
+  ttl_seconds: 2592000     # 30 days retention (for persistent backends)
+  async_writes: false      # Enable async writes for better performance
+
 semantic_cache:
   enabled: true
   backend_type: "memory"  # Options: "memory", "milvus", or "hybrid"
@@ -37,14 +46,9 @@ semantic_cache:
   # backend_config_path: "config/milvus.yaml" # Path to Milvus config
 
   # Embedding model for semantic similarity matching
-  # Options: "bert" (fast, 384-dim), "qwen3" (high quality, 1024-dim, 32K context), "gemma" (balanced, 768-dim, 8K context)
-  # Default: "bert" (fastest, lowest memory)
-  embedding_model: "bert"
-
-bert_model:
-  model_id: models/mom-embedding-light
-  threshold: 0.6
-  use_cpu: true
+  # If not specified, automatically uses the model configured in embedding_models section
+  # Options: "mmbert" (multilingual, 768-dim), "qwen3" (high quality, 1024-dim, 32K context), "gemma" (balanced, 768-dim, 8K context)
+  # embedding_model: "mmbert"  # Optional: explicitly set if you want to override auto-detection
 
 tools:
   enabled: true
@@ -125,20 +129,29 @@ feedback_detector:
 #     access_key: ""  # Optional: for Authorization header (Bearer token)
 
 # Embedding Models Configuration
-# These models provide intelligent embedding generation with automatic routing:
-# - Qwen3-Embedding-0.6B: Up to 32K context, high quality,
-# - EmbeddingGemma-300M: Up to 8K context, fast inference, Matryoshka support (768/512/256/128)
+# This is the UNIFIED configuration for all embedding-related features:
+# - Semantic Cache: Automatically uses the configured model
+# - Tool Selection: Uses the configured model for tool matching
+# - Embedding Signal: Uses the model specified in hnsw_config.model_type
+# - Complexity Signal: Uses the model specified in hnsw_config.model_type
+#
+# Available models:
+# - Qwen3-Embedding-0.6B (Pro): Up to 32K context, high quality, 1024-dim
+# - EmbeddingGemma-300M (Flash): Up to 8K context, fast inference, Matryoshka support (768/512/256/128)
+# - mmBERT-Embed-32K-2D-Matryoshka (Ultra): Up to 32K context, 1800+ languages, 2D Matryoshka (layer early exit + dimension reduction)
 embedding_models:
-  qwen3_model_path: "models/mom-embedding-pro"
+  # qwen3_model_path: "models/mom-embedding-pro"
   # gemma_model_path: "models/mom-embedding-flash"
+  mmbert_model_path: "models/mom-embedding-ultra"
   use_cpu: true  # Set to false for GPU acceleration (requires CUDA)
   # HNSW Configuration
   # Improves performance by preloading candidate embeddings at startup
   # and using HNSW index for O(log n) similarity search
   hnsw_config:
-    model_type: "qwen3"         # Which model to use: "qwen3" (high quality) or "gemma" (fast)
+    model_type: "mmbert"         # Which model to use: "qwen3" (high quality), "gemma" (fast), or "mmbert" (multilingual)
     preload_embeddings: true    # Precompute candidate embeddings at startup
-    target_dimension: 1024       # Embedding dimension
+    target_dimension: 768        # Embedding dimension (1024 for qwen3, 768 for gemma/mmbert)
+    # For mmbert only: target_layer (3/6/11/22) for layer early exit
     enable_soft_matching: true
     min_score_threshold: 0.5
 

{vllm_sr-0.1.0b2.dev20260126190945 → vllm_sr-0.1.0b2.dev20260202073049}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "vllm-sr"
-version = "0.1.0.beta.2.dev20260126190945"
+version = "0.1.0.beta.2.dev20260202073049"
 description = "vLLM Semantic Router - Intelligent routing for Mixture-of-Models"
 authors = [{name = "vLLM-SR Team"}]
 readme = "README.md"

{vllm_sr-0.1.0b2.dev20260126190945 → vllm_sr-0.1.0b2.dev20260202073049/vllm_sr.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: vllm-sr
-Version: 0.1.0b2.dev20260126190945
+Version: 0.1.0b2.dev20260202073049
 Summary: vLLM Semantic Router - Intelligent routing for Mixture-of-Models
 Author: vLLM-SR Team
 License: Apache-2.0