MaxKB vs FastGPT: Procurement vs Production Control

MaxKB focuses on a straightforward domestic private-delivery model. The offline installation path is clear. Professional and Enterprise authorization models, next-year maintenance, and vendor support tiers are public on the official website, with support divided into 5x8 and 7x24 levels. One-time authorization helps internal approval. The database query tool is clearly listed in official documentation, giving the product strong visibility. Projects that want "install and use, report the budget once" fit this product shape well.

FastGPT focuses on fine-grained control in complex production scenarios. Isolated code sandboxing, manual workflow pause and in-place recovery, unified quotas and usage governance, image vectors, multi-index retrieval, and cost explanation all point to one production question: when the system runs in real business for a long time, can problems be located, can humans intervene, and can costs be explained? FastGPT also keeps three paths: Cloud, community self-hosting, and managed / self-hosted commercial editions, allowing teams to validate at low cost before migrating to a private environment.

One-time authorization, maintenance ratio, and support tiers make budget approval more intuitive. Three-year total cost should also include hardware, model consumption, implementation person-days, cluster expansion, and feature customization. The three-year TCO table should include all of these costs.

2.1 FastGPT's Differentiated Capabilities

Production impact:

• For isolated sandboxes, security review should focus on network access, file access, process isolation, and audit traces. Runtime architecture determines the isolation boundary.
• Manual pause and in-place recovery lets AI enter approval workflows. The POC should record pause behavior, state persistence, and recovery path.
• Unified quotas control share-link and API Key usage. Cost split helps locate parsing, vectorization, retrieval, and generation cost.
• Image vectors and cost explanation cover chart and screenshot retrieval plus stage-level cost for each Q&A.

2.2 Shared Capabilities With Different Implementation Paths

Acceptance condition: For enterprise governance, fix organization, department, group, resource, log retention, and operation-type permission cases. Record multi-tenancy and cluster version boundaries in the POC.

2.3 Where MaxKB Has an Advantage

• Straightforward domestic private-delivery form: Offline installation path, authorization model, next-year maintenance, and vendor support tiers are public on the official website.
• Procurement predictability: One-time authorization and public support tiers simplify budget approval. Maintenance, implementation, models, and hardware still enter three-year TCO.
• Product visibility of the database query tool: The official documentation clearly lists a database query tool. FastGPT's general path mainly uses HTTP API and tool wrapping, which is more open and adds an integration layer.
• Public support tiers: Professional 5x8 and Enterprise 7x24 are more intuitive for buyers. Specific response and recovery targets should still be based on the contract.

3.1 Public Boundaries on Both Sides

GPLv3 and the FastGPT Open Source License create different constraint profiles: distribution, derivative works, multi-tenant SaaS operation, and de-branding boundaries should each enter the legal checklist. X-Pack commercial boundaries need written confirmation.

Internal enterprise use and external delivery should separately list secondary development ownership, distribution obligations, de-branding rights, and multi-customer service model, then check each LICENSE against the actual delivery path.

Authorization price, maintenance ratio, subscription price, and plan boundaries should use same-day written quotes from both sides. Record tax basis, authorization period, and implementation/maintenance scope together.

The procurement checklist should require a responsibility matrix, incident levels, recovery targets, and upgrade/rollback owners. Record first-response targets separately from repair time limits. For buyout models, additionally confirm whether next-year maintenance keeps the same service coverage. Final commitments should be based on the contract.

3.2 Three-Year TCO for Buyout and Subscription

Compare buyout and subscription under the same three-year scale. The cost table should record license, maintenance, hardware, model usage, implementation, cluster, upgrade, secondary development, and risk reserve:

The output is a three-year TCO table, plus sensitivity analysis for member count, document volume, request volume, and dual-node requirements.

The subscription-vs-buyout conclusion should be based on a same-scope three-year table. Record scale, growth assumptions, and internal budgeting model together.

In addition to the unified POC measurement table, add three test groups for sandboxing, human recovery, and quota cost:

Complete unified POC measurement table:

Put the following metrics into the "not publicly listed / POC required / contract required" checklist: steady-state and peak concurrency, requests per second, token throughput per minute, parsing and retrieval latency under specified hardware, maximum application and tenant counts, availability SLA and RTO / RPO, and real-time template and plugin counts.

Security, reliability, and accuracy should be confirmed through unified POC and contract acceptance. Controls that are not covered should enter the supplemental test checklist.