Multi-Model Architecture and Quota Multiplier Billing Practices for Smart Office Platforms

    This article explores the transition from a single-model to a multi-model architecture in smart office platforms. It details how to select appropriate models for various tasks, explains the quota multiplier billing mechanism, and provides best practices for enterprise quota management, cost optimization, and future multi-model orchestration.

    From Single to Multi-Model: Why Integrate Third-Party Models?

    In enterprise environments, tasks such as text refinement, formatting adjustments, short translations, long-text analysis, and code refactoring vary significantly in difficulty. It is challenging for a single model to simultaneously deliver high effectiveness, fast response times, and low costs across all these tasks.

    Previously, a smart office platform relied on two proprietary models: a lightweight version and a professional version. The lightweight model offered fast responses, making it suitable for drafting emails and meeting minutes. The professional model was adequate for daily tasks, but its capability boundaries became apparent in cross-domain knowledge Q&A and complex reasoning. The recent upgrade integrates multiple mainstream third-party models alongside the proprietary ones, allowing users to switch between them from the model selection list.

    When selecting a model, you can make an initial assessment based on the task type:

    Task Type Recommended Priority Key Metrics to Consider
    Formatting, short translation, key point extraction Lightweight proprietary model Quota consumption, response latency
    Email drafting, meeting minutes Lightweight or professional proprietary model Quota consumption, output quality
    Long-text deep analysis, cross-domain Q&A Third-party large model Output quality, failure rate
    Complex code refactoring Third-party large model Output quality, failure rate, retry count

    The division in the table is for reference only. Actual performance should be validated through small-scale testing with your own documents and scenarios.

    How Quota Consumption Multipliers Work

    In a multi-model environment, the underlying compute costs for different models to complete the same task can vary greatly. If all models display a single unified price, users cannot determine which model offers better value. Therefore, the platform indicates a quota consumption multiplier for each model on the selection page. A lower multiplier means fewer quotas are consumed for the same task.

    Underlying large models typically price input tokens and output tokens separately. The platform’s quota multiplier can be understood as a coefficient that converts underlying token consumption into a unified “AI points” system. The deduction for a single API call can be broken down into two parts:

    • Input part: Number of input tokens × Input unit price × Multiplier
    • Output part: Number of output tokens × Output unit price × Multiplier

    The following pseudocode illustrates the billing logic and does not represent the platform’s actual implementation:

    1. # Pseudocode for illustration only, does not represent actual platform implementation
    2. def estimate_quota(input_tokens, output_tokens, price_in, price_out, multiplier):
    3. input_cost = input_tokens * price_in
    4. output_cost = output_tokens * price_out
    5. return (input_cost + output_cost) * multiplier

    When using multiplier-based billing, several boundary conditions must be confirmed:

    • Whether the multiplier applies uniformly to both input and output.
    • Whether tokens that hit the cache are billed at a lower unit price or the same as regular tokens.
    • Whether quotas are deducted for failed, timed-out, or interrupted requests.
    • Whether automatic retries incur additional billing.

    These rules should be documented in the product documentation. Any adjustments to the multiplier must be announced with an effective date, and historical billing records should be retained for user verification.

    Unified Enterprise Quota Management

    When enterprises procure AI services, they typically go through an approval and compliance review process. The platform initially opened third-party model configurations to team and enterprise subscription versions, later expanding to individual and small team users.

    For enterprises, the most direct benefit is a shortened procurement process. Enterprises no longer need to separately sign contracts with various model providers, integrate APIs, or apply for individual quotas. All models are accessed through the platform’s unified API gateway, and usage is deducted uniformly from the enterprise account’s existing AI points.

    When deploying, you can prepare using the following checklist:

    1. Allocate quotas by department or project, and set usage limits for each account.
    2. Set different model usage permissions for administrators and regular members.
    3. Enable API logging to record the model used, timestamp, department, and deducted quota.
    4. Set alert thresholds, such as notifying the person in charge when monthly quota usage exceeds 80%.

    Time-Based Pricing and High-Frequency Scenario Optimization

    The platform has also introduced cost optimization strategies for all users, primarily falling into two categories.

    The first is time-based pricing. The load on AI computing power exhibits clear peak and off-peak periods throughout the day. The platform offers quota discounts for participating models during specified off-peak hours on weekdays and holidays. Users can utilize these services during off-peak times, allowing the platform to utilize its GPU clusters more evenly.

    The second is multiplier reductions for high-frequency, lightweight scenarios. The platform reduced the multipliers for two models oriented toward high-frequency, lightweight tasks. Tasks like formatting, short translations, and key point extraction are frequent in daily office work. While the cost per call is low, the cumulative volume accounts for a significant portion of usage. Following the multiplier reduction, the cost of these operations decreases proportionally. Actual savings depend on your task distribution and the magnitude of the multiplier adjustment; it is recommended to validate this using your own usage data.

    How can you determine if cost reduction is effective? You can track three metrics:

    • Quota consumption: Track the average deducted quota per task type on a weekly basis and compare the changes before and after the optimization.
    • Task latency: Record the time from submission to result return to confirm whether the lightweight model’s response speed meets expectations.
    • Failure rate: Track the number of failures and retries to ensure that failed requests are not billed multiple times.

    Recommendations for Users

    Start by using lightweight models for routine tasks, and only switch to third-party models for scenarios like long-text analysis or complex code refactoring. Before switching, check the multiplier annotations and compare output quality and quota consumption using a small sample. If you encounter billing discrepancies, you can cross-reference the API logs and the effective time of the multipliers.

    Multi-model orchestration may be the next step. The system could automatically select a model based on the complexity of the input task and set fallback strategies, such as switching to a backup model if one model times out. However, such mechanisms require clear orchestration rules, thresholds, and billing criteria, allowing users to clearly know which model executed each call and how much quota was consumed.