feat: LLM Router extension for cost-optimized model selection

redisvl/extensions/llm_router/DESIGN.md

@@ -0,0 +1,385 @@
+# LLM Router Extension - Design Document
+
+## Overview
+
+The LLM Router is an extension to RedisVL that provides intelligent, cost-optimized LLM model selection using semantic routing. Instead of routing queries to topics (like SemanticRouter), it routes queries to **model tiers** - selecting the cheapest LLM capable of handling each task.
+
+## Problem Statement
+
+### The LLM Cost Problem
+Modern applications often default to using the most capable (and expensive) LLM for all queries, even when simpler models would suffice:
+- "Hello, how are you?" -> Claude Opus 4.5 ($5/M tokens)
+- "Hello, how are you?" -> GPT-4.1 Nano ($0.10/M tokens)
+
+### Existing Solutions and Their Limitations
+
+**RouteLLM** (CMU/LMSys):
+- Binary classification only (strong vs weak model)
+- No support for >2 tiers
+- Requires training data or preference matrices
+
+**NVIDIA LLM Router Blueprint**:
+- Complexity classification approach (simple/moderate/complex)
+- Provides the taxonomy basis but no open-source Redis-native implementation
+
+**RouterArena / Bloom's Taxonomy Approach**:
+- Maps query complexity to Bloom's cognitive levels
+- Informs our tier design but lacks production routing infrastructure
+
+**OpenRouter Auto-Router**:
+- Black box routing decisions
+- Data flows through third-party servers
+- No transparency into why a model was selected
+- Can't self-host or customize
+
+**NotDiamond**:
+- Proprietary ML model for routing
+- Requires API calls for every routing decision
+- No local/offline capability
+
+**FrugalGPT**:
+- Sequential cascade approach (try cheap first, escalate)
+- Higher latency due to serial model calls
+
+## Solution: Semantic Model Tier Routing
+
+Repurpose RedisVL's battle-tested SemanticRouter for model selection:
+
+```
+SemanticRouter          ->    LLMRouter
+-----------------------------------------
+Route                   ->    ModelTier
+route.name              ->    tier.name (simple/standard/expert)
+route.references        ->    tier.references (task complexity examples)
+route.metadata          ->    tier.metadata (cost, capabilities)
+RouteMatch              ->    LLMRouteMatch (includes model string)
+```
+
+### Architecture
+
+```
++---------------------------------------------------------------+
+|                         LLMRouter                             |
++---------------------------------------------------------------+
+|  +-------------+  +-------------+  +-------------+           |
+|  |   Simple    |  |  Standard   |  |   Expert    |           |
+|  |   Tier      |  |    Tier     |  |    Tier     |           |
+|  +-------------+  +-------------+  +-------------+           |
+|  | gpt-4.1-nano|  | sonnet 4.5  |  | opus 4.5    |           |
+|  | $0.10/M     |  | $3/M        |  | $5/M        |           |
+|  | threshold:  |  | threshold:  |  | threshold:  |           |
+|  | 0.5         |  | 0.6         |  | 0.7         |           |
+|  +-------------+  +-------------+  +-------------+           |
+|         |                |                |                   |
+|         +----------------+----------------+                   |
+|                          v                                    |
+|              +------------------------+                       |
+|              |   Redis Vector Index   |                       |
+|              |   (reference phrases)  |                       |
+|              +------------------------+                       |
++---------------------------------------------------------------+
+                           |
+                           v
+                    +-------------+
+                    |   Query     |
+                    |  "analyze   |
+                    |   this..."  |
+                    +-------------+
+                           |
+                           v
+                    +-------------+
+                    |  LiteLLM    |
+                    |  (optional) |
+                    +-------------+
+```
+
+## Key Design Decisions
+
+### 1. Model Tiers, Not Individual Models
+
+Routes map to **tiers** (simple, standard, expert) rather than specific models. This provides:
+- Abstraction from model churn (swap haiku -> gemini-flash without changing routes)
+- Clear mental model for users
+- Easy cost optimization within tiers
+
+### 2. Bloom's Taxonomy-Grounded Tiers
+
+The default pretrained config maps tiers to Bloom's Taxonomy cognitive levels:
+- **Simple** (Remember/Understand): Factual recall, greetings, format conversion
+- **Standard** (Apply/Analyze): Code explanation, summarization, moderate analysis
+- **Expert** (Evaluate/Create): Research, architecture, formal reasoning
+
+This is informed by RouterArena's finding that cognitive complexity correlates with model capability requirements.
+
+### 3. LiteLLM-Compatible Model Strings
+
+Tier model identifiers use LiteLLM format (`provider/model`):
+```python
+ModelTier(
+    name="standard",
+    model="anthropic/claude-sonnet-4-5",  # Works directly with LiteLLM
+    ...
+)
+```
+
+### 4. Per-Tier Distance Thresholds
+
+Each tier has its own `distance_threshold`, allowing fine-grained control:
+```python
+simple_tier = ModelTier(..., distance_threshold=0.5)   # Strict match
+expert_tier = ModelTier(..., distance_threshold=0.7)   # Looser match
+```
+
+### 5. Cost-Aware Routing
+
+When `cost_optimization=True`, the router adds a cost penalty to distances:
+```python
+adjusted_distance = distance + (cost_per_1k * cost_weight)
+```
+This prefers cheaper tiers when semantic distances are close.
+
+### 6. Pretrained Configs with Embedded Vectors
+
+The built-in `default.json` provides a ready-to-use 3-tier configuration:
+```python
+# Instant setup - no embedding model needed at load time
+router = LLMRouter.from_pretrained("default", redis_client=client)
+```
+
+The pretrained config includes pre-computed embeddings from
+`sentence-transformers/all-mpnet-base-v2`, with 18 reference phrases per tier
+covering the Bloom's Taxonomy spectrum.
+
+Custom configs can also be exported and shared:
+```python
+# Export (one-time, with embedding model)
+router.export_with_embeddings("my_router.json")
+
+# Import (no embedding needed)
+router = LLMRouter.from_pretrained("my_router.json", redis_client=client)
+```
+
+### 7. Async Support
+
+`AsyncLLMRouter` provides the same functionality using async I/O. Since
+`__init__` cannot be async, it uses a `create()` classmethod factory:
+
+```python
+router = await AsyncLLMRouter.create(
+    name="my-router",
+    tiers=tiers,
+    redis_client=async_client,
+)
+match = await router.route("hello")
+```
+
+Key async method mapping:
+
+| Sync (`LLMRouter`) | Async (`AsyncLLMRouter`) |
+|---------------------|--------------------------|
+| `__init__()` | `await create()` |
+| `from_existing()` | `await from_existing()` |
+| `route()` | `await route()` |
+| `route_many()` | `await route_many()` |
+| `add_tier()` | `await add_tier()` |
+| `remove_tier()` | `await remove_tier()` |
+| `from_dict()` | `await from_dict()` |
+| `from_pretrained()` | `await from_pretrained()` |
+| `delete()` | `await delete()` |
+
+## Module Structure
+
+```
+redisvl/extensions/llm_router/
++-- __init__.py              # Public exports (LLMRouter, AsyncLLMRouter, schemas)
++-- DESIGN.md                # This document
++-- schema.py                # Pydantic models
+|   +-- ModelTier            # Tier definition
+|   +-- LLMRouteMatch        # Routing result
+|   +-- RoutingConfig        # Router configuration
+|   +-- Pretrained*          # Export/import schemas
++-- router.py                # LLMRouter + AsyncLLMRouter implementations
++-- pretrained/
+    +-- __init__.py          # Pretrained loader (get_pretrained_path)
+    +-- default.json         # Standard 3-tier config (simple/standard/expert)
+```
+
+## API Examples
+
+### Basic Usage
+
+```python
+from redisvl.extensions.llm_router import LLMRouter, ModelTier
+
+tiers = [
+    ModelTier(
+        name="simple",
+        model="openai/gpt-4.1-nano",
+        references=[
+            "hello", "hi there", "thanks", "goodbye",
+            "what time is it?", "how are you?",
+        ],
+        metadata={"cost_per_1k_input": 0.0001},
+        distance_threshold=0.5,
+    ),
+    ModelTier(
+        name="standard",
+        model="anthropic/claude-sonnet-4-5",
+        references=[
+            "analyze this code for bugs",
+            "explain how neural networks learn",
+            "compare and contrast these approaches",
+        ],
+        metadata={"cost_per_1k_input": 0.003},
+        distance_threshold=0.6,
+    ),
+    ModelTier(
+        name="expert",
+        model="anthropic/claude-opus-4-5",
+        references=[
+            "prove this mathematical theorem",
+            "architect a distributed system",
+            "write a research paper analyzing",
+        ],
+        metadata={"cost_per_1k_input": 0.005},
+        distance_threshold=0.7,
+    ),
+]
+
+router = LLMRouter(
+    name="my-llm-router",
+    tiers=tiers,
+    redis_url="redis://localhost:6379",
+)
+
+# Route a query
+match = router.route("hello, how's it going?")
+print(match.tier)   # "simple"
+print(match.model)  # "openai/gpt-4.1-nano"
+
+# Use with LiteLLM (optional integration)
+from litellm import completion
+response = completion(model=match.model, messages=[{"role": "user", "content": query}])
+```
+
+### Cost-Optimized Routing
+
+```python
+router = LLMRouter(
+    name="cost-aware-router",
+    tiers=tiers,
+    cost_optimization=True,  # Prefer cheaper tiers when distances are close
+    redis_url="redis://localhost:6379",
+)
+```
+
+### Pretrained Router
+
+```python
+# Load without needing an embedding model for the references
+router = LLMRouter.from_pretrained(
+    "default",  # Built-in config, or path to JSON
+    redis_client=client,
+)
+```
+
+### Async Usage
+
+```python
+from redisvl.extensions.llm_router import AsyncLLMRouter
+
+router = await AsyncLLMRouter.create(
+    name="my-async-router",
+    tiers=tiers,
+    redis_url="redis://localhost:6379",
+)
+
+match = await router.route("explain how garbage collection works")
+print(match.model)  # "anthropic/claude-sonnet-4-5"
+
+# Or load from pretrained
+router = await AsyncLLMRouter.from_pretrained("default", redis_client=client)
+
+await router.delete()
+```
+
+## Comparison with SemanticRouter
+
+| Feature | SemanticRouter | LLMRouter |
+|---------|---------------|-----------|
+| Purpose | Topic classification | Model selection |
+| Output | Route name | Model string + metadata |
+| Cost awareness | No | Yes |
+| Pretrained configs | No | Yes |
+| Per-route thresholds | Yes | Yes |
+| LiteLLM integration | No | Yes (model strings) |
+| Async support | No | Yes (`AsyncLLMRouter`) |
+
+## Testing
+
+### Unit Tests (`tests/unit/test_llm_router_schema.py`)
+- Schema validation
+- Pydantic model behavior
+- Threshold bounds
+- Empty/invalid inputs
+
+### Integration Tests (`tests/integration/test_llm_router.py`)
+- Router initialization
+- Routing accuracy
+- Cost optimization behavior
+- Serialization (dict, YAML, JSON)
+- Pretrained import/export
+- Pretrained config loading (`from_pretrained("default")`)
+- Tier management (add, remove, update)
+- Persistence (from_existing)
+
+### Async Integration Tests (`tests/integration/test_async_llm_router.py`)
+- Mirrors all sync tests with `AsyncLLMRouter`
+- Uses `async_client` fixture and async skip helpers
+- Tests `create()` factory, async routing, serialization, tier management
+- Pretrained config loading
+
+Run tests:
+```bash
+uv run pytest tests/unit/test_llm_router_schema.py -v
+uv run pytest tests/integration/test_llm_router.py -v
+uv run pytest tests/integration/test_async_llm_router.py -v
+```
+
+## Future Enhancements
+
+### 1. `complete()` Method
+Direct LiteLLM integration for one-liner usage:
+```python
+response = router.complete("analyze this code", messages=[...])
+```
+
+### 2. Capability Filtering
+Filter tiers by capability before routing:
+```python
+match = router.route("generate an image", capabilities=["vision"])
+```
+
+### 3. Budget Constraints
+Enforce cost limits:
+```python
+router = LLMRouter(..., max_cost_per_1k=0.01)  # Never select opus
+```
+
+### 4. Fallback Chains
+Define fallback order when primary tier unavailable:
+```python
+tier = ModelTier(..., fallback=["standard", "simple"])
+```
+
+## References
+
+- [RedisVL SemanticRouter](https://docs.redisvl.com/en/latest/user_guide/semantic_router.html)
+- [LiteLLM Model List](https://docs.litellm.ai/docs/providers)
+- [RouteLLM](https://github.com/lm-sys/RouteLLM) - LMSys binary router framework
+- [NVIDIA LLM Router Blueprint](https://build.nvidia.com/blueprints/llm-router) - Complexity-based routing
+- [RouterArena / Bloom's Taxonomy](https://arxiv.org/abs/2412.06644) - Cognitive complexity for routing
+- [FrugalGPT](https://arxiv.org/abs/2305.05176) - Cost-efficient LLM strategies
+- [OpenRouter](https://openrouter.ai/) - Auto-routing concept
+- [NotDiamond](https://notdiamond.ai/) - ML-based model routing
+- [Unify.ai](https://unify.ai/) - Quality-cost tradeoff routing