Files
gpt4free/docs/config-yaml-routing.md
T
2026-03-19 15:28:00 +00:00

6.7 KiB
Raw Blame History

Custom Model Routing with config.yaml

g4f supports a config.yaml file that lets you define custom model routes named models that are transparently forwarded to one or more real providers based on availability, quota balance, and recent error counts.

This is similar to the LiteLLM routing configuration.


Quick start

  1. Place a config.yaml file in the same directory as your .har / .json cookie files (the "cookies dir").

    • Default location: ~/.config/g4f/cookies/config.yaml
    • Alternative: ./har_and_cookies/config.yaml
  2. Define your routes (see format below).

  3. g4f loads the file automatically when it reads the cookie directory (e.g. on API server start-up, or when read_cookie_files() is called).

  4. Request the custom model name from any client:

from g4f.client import Client

client = Client()
response = client.chat.completions.create(
    model="my-gpt4",   # defined in config.yaml
    messages=[{"role": "user", "content": "Hello!"}],
)
print(response.choices[0].message.content)

File format

models:
  - name: "<model-name>"          # the name clients use
    providers:
      - provider: "<ProviderName>"  # g4f provider class name
        model: "<provider-model>"   # model name passed to that provider
        condition: "<expression>"   # optional  see below
      - provider: "..."             # fallback provider (no condition = always eligible)
        model: "..."

Keys

Key Required Description
name The model name used by clients.
providers Ordered list of provider candidates.
provider Provider class name (e.g. "OpenaiAccount", "PollinationsAI").
model Model name forwarded to the provider. Defaults to the route name.
condition Boolean expression controlling when this provider is eligible.

Condition expressions

The condition field is a boolean expression evaluated before each request. It can reference the following variables:

quota full provider quota dict

Each provider that implements get_quota() returns a provider-specific dict. The result is cached in memory (5 min TTL) and invalidated on 429 responses.

Access any field with dot-notation:

Provider get_quota() format Example condition
PollinationsAI {"balance": float} quota.balance > 0
Yupp {"credits": {"remaining": int, "total": int}} quota.credits.remaining > 100
PuterJS raw metering JSON from the API quota.total_requests < 1000
GeminiCLI {"buckets": [...]} error_count < 3
GithubCopilot usage details dict error_count < 5

Missing keys resolve to 0.0 (no error raised).

balance shorthand alias

balance is a convenience shorthand for quota.balance. It is preserved for backward compatibility and is most useful with PollinationsAI which returns {"balance": float}. For other providers, prefer the explicit quota.* form.

error_count

Number of errors recorded for this provider in the last 1 hour. Errors older than 1 hour are automatically pruned.

Operators

Operator Meaning
> < >= <= Numeric comparison
== != Equality / inequality
and or not Logical connectives
( ) Grouping

Examples

# PollinationsAI  uses quota.balance shorthand
condition: "balance > 0"
condition: "balance > 0 or error_count < 3"

# Yupp  provider-specific nested field
condition: "quota.credits.remaining > 0"
condition: "quota.credits.remaining > 0 or error_count < 3"

# Any provider  error-count-only conditions work universally
condition: "error_count < 3"
condition: "error_count == 0"

When the condition is absent or evaluates to True, the provider is eligible. When it evaluates to False the provider is skipped and g4f tries the next one in the list.


Quota caching

Quota values are fetched via the provider's get_quota() method and cached in memory for 5 minutes (configurable via QuotaCache.ttl).

When a provider returns an HTTP 429 (Too Many Requests) error the cache entry for that provider is immediately invalidated, so the next routing decision fetches a fresh quota value before deciding.


Error counting

Every time a provider raises an exception the error counter for that provider is incremented. Errors older than 1 hour are automatically pruned.

Reference error_count in a condition to avoid retrying providers that have been failing repeatedly.


Full example

# ~/.config/g4f/cookies/config.yaml

models:
  # PollinationsAI: use quota.balance shorthand
  - name: "my-gpt4"
    providers:
      - provider: "OpenaiAccount"
        model: "gpt-4o"
        condition: "balance > 0 or error_count < 3"
      - provider: "PollinationsAI"
        model: "openai-large"

  # Yupp: provider-specific nested quota field
  - name: "yupp-chat"
    providers:
      - provider: "Yupp"
        model: "gpt-4o"
        condition: "quota.credits.remaining > 0 or error_count < 3"
      - provider: "PollinationsAI"
        model: "openai-large"

  # Universal: error-count-only condition works for any provider
  - name: "llama-fast"
    providers:
      - provider: "Groq"
        model: "llama-3.3-70b"
        condition: "error_count < 3"
      - provider: "DeepInfra"
        model: "meta-llama/Llama-3.3-70B-Instruct"

Python API

The routing machinery is exposed in g4f.providers.config_provider:

from g4f.providers.config_provider import (
    RouterConfig,        # load / query routes
    QuotaCache,          # inspect / invalidate quota cache
    ErrorCounter,        # inspect / reset error counters
    evaluate_condition,  # evaluate a condition string directly
)

# Reload routes from a custom path
RouterConfig.load("/path/to/config.yaml")

# Check if a route exists
route = RouterConfig.get("my-gpt4")  # returns ModelRouteConfig or None

# Manually invalidate quota cache (e.g. after detecting 429)
QuotaCache.invalidate("OpenaiAccount")

# Check error count
count = ErrorCounter.get_count("OpenaiAccount")

# Evaluate a condition string with a full provider-specific quota dict
# (PollinationsAI)
ok = evaluate_condition("balance > 0 or error_count < 3", {"balance": 0.0}, 2)
# True

# Yupp-style nested quota
ok = evaluate_condition(
    "quota.credits.remaining > 0",
    {"credits": {"remaining": 500, "total": 5000}},
    0,
)
# True

Requirements

PyYAML must be installed:

pip install pyyaml

It is included in the full requirements.txt. If PyYAML is absent g4f logs a warning and skips config.yaml loading.