A routing/orchestration library (adjust as needed)
Project description
PolyRouter
PolyRouter is a lightweight Python library that routes requests across multiple LLM providers. It helps applications achieve deterministic failover by rotating API keys, client providers, and model candidates when requests fail.
Overview
This repository provides an orchestration layer that can be embedded in your application to manage provider rotation, API-key pools, and model fallbacks. Keys are loaded from the environment (see .env), and the orchestrator tries configured provider/model/key combinations until a request succeeds or all combinations are exhausted.
Example provider adapters included in this snapshot:
- Groq
- Google Gemini
- Cerebras
Badges
| Badge | Meaning |
|---|---|
| Python implementation | |
| Multi-provider failover design | |
| Update once a formal license is chosen |
Key Features
| Capability | Technical Detail |
|---|---|
| Provider rotation | Requests can move across Groq, Gemini, and Cerebras client pools without changing application code. |
| Key pool management | Each provider can be backed by multiple API keys, allowing the runtime to continue when a single key expires or is rate-limited. |
| Model pool fallback | Ordered model lists in config.py act as a preference chain, so the router can try alternate models before surfacing a failure. |
| Debug visibility | DEBUG and IN_DEPTH_DEBUG control log verbosity so you can switch between concise operational logs and deep request tracing. |
| Centralized configuration | Provider counts, model lists, and debug mode live in one place instead of being duplicated across call sites. |
| Failure isolation | Provider-specific errors do not have to terminate the entire workflow if another valid key/model combination is still available. |
How It Works
flowchart TD
A[Application request] --> B[Load config.py]
B --> C[Load .env API keys]
C --> D[Try primary provider]
D --> E{Request succeeds?}
E -->|Yes| F[Return response]
E -->|No| G[Rotate key / model / client]
G --> H{Any combinations left?}
H -->|Yes| D
H -->|No| I[Raise exhaustion error]
The intended behavior is simple:
- Read provider preferences, model lists, and key counts from
config.py. - Load API credentials from the environment.
- Attempt a request with the active provider/model combination.
- On provider failure, rotate through the next key or model.
- When a provider pool is exhausted, move to the next client family.
- Stop only when every configured combination has been tried.
The repo is built for operational resilience, not for single-provider purity.
Project Structure
PolyRouter/
├── examples/ # Example usage scenarios
│ └── basic_usage.py
├── polyrouter/ # Library source
│ ├── Exceptions.py
│ ├── LLMClients.py
│ ├── LLMOrchestrator.py
│ └── __init__.py
├── .env # (example present in repo snapshot)
├── requirements-dev.txt # Development / runtime deps
└── README.md
Installation
Local setup (recommended):
git clone <repository-url>
cd PolyRouter
python3 -m venv .venv
source .venv/bin/activate
pip install --upgrade pip
pip install -r requirements-dev.txt
Environment variables
This repo includes a .env file in the snapshot. In normal usage copy and populate a local .env (do not commit secrets):
cp .env .env.local
# edit .env.local and export provider API keys, e.g.:
# GROQ_API_KEY0=...
# GEMINI_API_KEY0=...
The examples use environment variables named like GROQ_API_KEY0, GEMINI_API_KEY0, etc.
Dependencies
Dependencies used by examples and adapters may include provider SDKs and python-dotenv. Install via requirements-dev.txt.
Build / verification steps
This repository is a Python library-style project rather than a packaged service, so the essential validation step is import verification plus a smoke test in your host application.
python -m compileall .
python - <<'PY'
from config import DEBUG, IN_DEPTH_DEBUG, GROQ_MODEL
print("config loaded:", DEBUG, IN_DEPTH_DEBUG, GROQ_MODEL)
PY
Configuration
config.py is the primary customization point.
| Setting | Role |
|---|---|
DEBUG |
Enables the main debug statement stream. |
IN_DEPTH_DEBUG |
Enables detailed trace output for low-level troubleshooting. |
GROQ_MODEL |
Ordered Groq model preference list. |
GEMINI_MODEL |
Ordered Gemini model preference list. |
CEREBRAS_MODEL |
Ordered Cerebras model preference list. |
GROQ_KEY |
Number of Groq API keys to scan. |
GEMINI_KEY |
Number of Gemini API keys to scan. |
CEREBRAS_KEY |
Number of Cerebras API keys to scan. |
Recommended operating model:
DEBUG = 1
IN_DEPTH_DEBUG = 0
GROQ_MODEL = ["openai/gpt-oss-120b", "llama-3.3-70b-versatile"]
GEMINI_MODEL = ["gemini-2.5-flash"]
CEREBRAS_MODEL = ["gpt-oss-120b"]
GROQ_KEY = 2
GEMINI_KEY = 1
CEREBRAS_KEY = 1
Usage
See examples/basic_usage.py for a minimal example. The orchestrator can be constructed directly in your application; you do not need a config.py file if you prefer to pass provider settings programmatically.
Minimal usage:
from dotenv import load_dotenv
from polyrouter.LLMOrchestrator import LLMOrchestrator
import os
load_dotenv()
llm = LLMOrchestrator(
groq={
"groq_models": ["openai/gpt-oss-120b"],
"groq_keys": [os.getenv("GROQ_API_KEY0")],
},
debug=True,
verbose=True,
prompt="You are a helpful assistant",
)
response = llm.request()
print(response)
When debug/verbose are enabled the orchestrator prints provider, model and key selection and rotation decisions.
API / CLI
No standalone CLI or HTTP API is exposed in this repository snapshot.
The public surface is intentionally library-oriented:
config.pycontrols behaviorLLMClients.pydefines the client abstractionLLMOrchestrator.pyis the orchestration boundary
If you add a CLI later, document it here with exact command syntax and exit codes.
Deployment
Because this project is a routing library, deployment usually means shipping it as part of a larger Python service or worker.
Recommended deployment checklist
- Pin dependencies with
requirements.txt. - Inject secrets through the runtime environment, not source control.
- Set
DEBUG = 0andIN_DEPTH_DEBUG = 0for production unless you are actively diagnosing issues. - Validate all required API keys are present before starting the process.
- Run the host service behind your preferred process manager, container runtime, or platform scheduler.
Containerized deployment
If you package the project into a container, copy only the source files, install requirements, and mount secrets through environment variables or secret storage.
FROM python:3.11-slim
WORKDIR /app
COPY requirements.txt .
RUN pip install --no-cache-dir -r requirements.txt
COPY . .
CMD ["python", "-c", "import config; print('LLM-Gateway-Service ready')"]
Screenshots
Screenshot placeholder: add an architecture diagram or runtime trace capture here once the project has a visual demo surface.
Suggested assets for a production repository:
- request-routing diagram
- provider rotation log snippet
- environment setup screenshot
Troubleshooting
| Symptom | Likely Cause | Resolution |
|---|---|---|
| Import error for a provider SDK | Dependencies are missing from the active virtual environment | Re-run pip install -r requirements.txt inside the activated environment. |
| Requests stop after one provider fails | No fallback keys or models are configured | Add more keys to .env and expand the model pool in config.py. |
| All requests fail immediately | Environment variables are missing or misnamed | Verify .env matches .env.template exactly. |
| Debug logs are too noisy | Verbosity flags are enabled | Set DEBUG = 0 and IN_DEPTH_DEBUG = 0 for normal operation. |
| A specific model keeps failing | The model is unsupported, rate-limited, or exhausted | Remove it from the preference list or move it later in the rotation order. |
Contributing
Contributions are welcome if they improve correctness, observability, or provider coverage.
Please keep pull requests focused and include:
- a concise description of the routing behavior being changed
- reproduction steps for any failure-handling update
- updates to
.env.templateandconfig.pywhen configuration contracts change - tests or a clear validation checklist when the orchestration flow changes
Guidelines:
- Do not hard-code secrets.
- Preserve backward-compatible configuration names whenever possible.
- Keep provider rotation behavior deterministic and well logged.
- Prefer small, isolated changes to client adapters and error handling.
Roadmap
Planned improvements that would strengthen the project further:
- formal public orchestration API with documented inputs and return types
- structured logging with request IDs and provider attempt history
- health checks for provider pools and exhausted key detection
- test coverage for failover, invalid-key handling, and model rotation
- optional CLI for smoke testing provider credentials
- metrics hooks for success rate, fallback rate, and exhaustion rate
Acknowledgements
LLM-Gateway-Service builds on the ecosystem provided by:
- Groq
- Google Gemini / Google Gen AI SDK
- Cerebras Cloud SDK
- python-dotenv
- tenacity
It also follows a common open-source reliability pattern: fail over without forcing callers to understand vendor-specific error recovery.
License
License: TBD.
Add the repository's chosen license here once it is finalized, and keep the license file in sync with this section.
Project details
Release history Release notifications | RSS feed
Download files
Download the file for your platform. If you're not sure which to choose, learn more about installing packages.
Source Distribution
Built Distribution
Filter files by name, interpreter, ABI, and platform.
If you're not sure about the file name format, learn more about wheel file names.
Copy a direct link to the current filters
File details
Details for the file polyrouter-1.0.0.tar.gz.
File metadata
- Download URL: polyrouter-1.0.0.tar.gz
- Upload date:
- Size: 15.3 kB
- Tags: Source
- Uploaded using Trusted Publishing? No
- Uploaded via: twine/6.2.0 CPython/3.11.8
File hashes
| Algorithm | Hash digest | |
|---|---|---|
| SHA256 |
67f40aee5efdf47bc51b27a34de653c760783c54e8622d3dcb3300d77f511d29
|
|
| MD5 |
76b9cd8a8532a7a62966f3f09d256fe6
|
|
| BLAKE2b-256 |
0c8a07c6995a6d53d0cfd2e570dea103432e7b4a0ad961d6c06932e6dabecd26
|
File details
Details for the file polyrouter-1.0.0-py3-none-any.whl.
File metadata
- Download URL: polyrouter-1.0.0-py3-none-any.whl
- Upload date:
- Size: 11.8 kB
- Tags: Python 3
- Uploaded using Trusted Publishing? No
- Uploaded via: twine/6.2.0 CPython/3.11.8
File hashes
| Algorithm | Hash digest | |
|---|---|---|
| SHA256 |
c0450989ad615cdc09342abc0a5293a0b8a7cd7509ec7cae571ba68850b1b1d8
|
|
| MD5 |
cf3667596c0567a1b24c6c1e09af4e92
|
|
| BLAKE2b-256 |
43e951bc2c035e884260e1ae888b0e79507fa10f1ba9cca839df8ff08976b123
|