synaptoroute 0.3.1

A local, blazingly fast Semantic Router designed for Agentic AI workflows.

SynaptoRoute

SynaptoRoute is a high-performance semantic routing engine designed for production Python microservices. It executes intent classification locally using hardware-accelerated vector embeddings to route natural language to deterministic software logic.

It is specifically designed for:

• Tool Routing (Mapping prompts to function signatures)
• Agent Routing (Handing off state to specialized subagents)
• Workflow Orchestration (Triggering RAG chains or DB queries)
• Large-Scale Intent Classification (Supporting massive, dense domain definitions)

Key Features

✓ Async Batching Queue: Prevents event loop blocking and absorbs massive concurrent loads without hardware lockup. ✓ Fast Route Mutations: Hot-swap intents in near-constant-time memory without rebuilding the search index. ✓ 50,000+ Route Support: Backed by an optional Faiss index to maintain interactivity at massive scale. ✓ Pluggable Architecture: Seamlessly swap embedding providers (Local ONNX, OpenAI, etc.). ✓ Distributed Sync: Redis-backed pub/sub to keep Kubernetes replicas aligned.

---

Benchmark Highlights

SynaptoRoute has been rigorously benchmarked against both procedural stressors and non-synthetic canonical datasets.

• 50,000 routes tested interactively.
• ~49ms retrieval latency at 250,000 embeddings.
• CLINC150: 74.20% Top-1 Accuracy
• Banking77: 91.81% Top-1 Accuracy (Zero-shot mapping on 77 highly overlapping intents)
• ~302 QPS on concurrent batched benchmark workloads.

For full statistical breakdowns, methodology, and comparisons, see docs/BENCHMARKS.md and docs/COMPARISON.md.

---

When to Use SynaptoRoute

Use SynaptoRoute if:

• You need local, edge-deployed routing without API dependencies.
• You need high concurrency capable of surviving asynchronous spikes.
• You expect massive routing tables (1,000 to 50,000+ routes).
• You want highly predictable query latency regardless of scale.

Consider alternatives if:

• You need logical reasoning or downstream multi-step planning.
• You need complex multi-intent decomposition.
• You require strict Out-Of-Distribution detection without manual calibration.

---

Architecture & Design

In modern microservice architectures, relying on external APIs for classification routing introduces high latency, cost, and rate limits. SynaptoRoute executes intent classification locally, avoiding two structural bottlenecks common in semantic routing:

1. Sequential Compute Starvation: Processing single semantic requests sequentially creates a bottleneck for parallel API calls, eventually forcing thermal throttling or thread exhaustion on local hardware. SynaptoRoute captures concurrent requests in a background _batch_worker queue, groups them (e.g., batch size 32), and executes them in a single optimized pass through the inference engine.
2. Index Rebuilding Penalty: Standard routers execute an $O(N)$ reallocation of the entire memory space when routes change. SynaptoRoute utilizes lazy slicing and memory-mapped tombstoning to allow instant insertions and deletions.

---

1. Installation

pip install synaptoroute

# Optional Extras
pip install synaptoroute[api]          # For FastAPI integration
pip install synaptoroute[openai]       # For using OpenAI embeddings
pip install synaptoroute[metrics]      # For telemetry and evaluation
pip install synaptoroute[redis]        # For distributed deployment syncing
pip install synaptoroute[faiss]        # For massive route scaling (50,000+)
pip install synaptoroute[langchain]    # For LangChain ecosystem integration
pip install synaptoroute[llamaindex]   # For LlamaIndex ecosystem integration
pip install synaptoroute[all]          # Installs all optional dependencies

---

Quick Start Guide

Basic Example

import asyncio
from synaptoroute.router import AdaptiveRouter
from synaptoroute.encoder import FastEmbedEncoder
from synaptoroute.storage import SQLiteStorage
from synaptoroute.models import Route

async def main():
    # 1. Initialize Components
    encoder = FastEmbedEncoder(model_name="BAAI/bge-small-en-v1.5")
    storage = SQLiteStorage("data/memory.sqlite")
    router = AdaptiveRouter(encoder=encoder, storage=storage)
    
    # 2. Define Routes
    billing_route = Route(
        name="billing", 
        utterances=["I need a refund", "Where is my receipt?", "Cancel my subscription"],
        threshold=0.60
    )
    router.add_route(billing_route)
    
    # 3. Start the Background Batching Worker
    await router.start()
    
    # 4. Execute Async Queries
    result = await router.aquery("How do I get my money back?")
    if result:
        print(f"Matched Intent: {result.name}") # Output: billing
    
    # 5. Graceful Shutdown
    await router.stop()

if __name__ == "__main__":
    asyncio.run(main())

---

Advanced Configuration

Optimization Profiles

SynaptoRoute allows you to load strict optimization profiles depending on your infrastructure constraints:

from synaptoroute import OptimizationProfile, AdaptiveRouter

# THROUGHPUT: Maximizes QPS for heavy concurrent loads
router = AdaptiveRouter(profile=OptimizationProfile.THROUGHPUT)

# LATENCY: Minimizes response time for sequential or low-concurrency systems
router = AdaptiveRouter(profile=OptimizationProfile.LATENCY)

Caveat: profile.threads must be passed explicitly to FastEmbedEncoder. The router does not propagate thread count automatically.

Distributed Deployment

For multi-pod Kubernetes or horizontal scaling, SynaptoRoute uses RedisSyncManager to synchronize SQLite route databases across nodes:

from synaptoroute.sync import RedisSyncManager

sync_manager = RedisSyncManager(redis_url="redis://localhost:6379")
router = AdaptiveRouter(sync_manager=sync_manager)

Caveat: The current RedisSyncManager implementation does not retry on Redis disconnects.

Roadmap

SynaptoRoute is rapidly evolving from a high-speed text semantic router into a multi-modal, highly scalable routing framework for distributed architectures.

• v0.4.0 (Observability): Native OpenTelemetry/Prometheus /metrics integration to track P50/P99 latency, queue depth, and throughput in production.
• v0.5.0 (Dynamic Boundaries): Automatic docstring extraction and LLM-assisted synthetic utterance generation to seed intents with zero manual configuration.
• v0.6.0 (Cross-Encoder Reranking): Optional two-stage retrieval. When the top two routes score within a narrow margin, a Cross-Encoder (e.g., MS MARCO) will execute a surgical tie-break to improve lexical keyword traps and adversarial routing performance.
• v0.7.0 (Multi-Modal): CLIP/ImageBind integration to accept PIL.Image objects and route visual data directly to specialized subsystems.

For the detailed strategic vision, see docs/ROADMAP.md.

Known Limitations

1. Directional Semantics: Vector similarity cannot distinguish between "flight book" and "cancel flight".
2. Deep Logical Negation: Modifiers like "don't", "never", and "not" are inherently problematic for dense embeddings.
3. Threshold Calibration: Defining a global threshold across highly disparate intents requires manual tuning.
4. Mixed Intent Parsing: Cannot natively split multi-action sentences into discrete routes.
5. Context Amnesia: Evaluates single utterances strictly without conversation history.
6. Cross-Language Drift: Cosine boundary margins differ significantly when evaluating multiple languages simultaneously.
7. Adversarial Resilience: Keyword traps will natively bypass standard embeddings unless explicitly trained out.

For a detailed analysis of these failure modes and how to implement recommended fallback mechanisms (like LLM verification), please read our limitations documentation.

---

Community & Contributing

We welcome professional contributions to expand the framework.

• Contributing: Review the Contributing Guidelines for architectural enforcement policies.
• Issues: Report bugs or request features via the Issue Tracker.