Sign in

The platform normalizes model text across providers by stripping structural artifacts such as thinking blocks, numbered prefixes, and formatting delimiters, while deploying a fault-tolerant salvage parser to recover structured JSON data from truncated responses.

flowchart TD
    A[Gemini Thought Tags] --> D[Regex Cleaner]
    B[Anthropic Tool-Use JSON] --> E[Markdown Stripper]
    C[Truncated JSON Arrays] --> F[Salvage Parser]

    D --> G[Normalized Text Stream]
    E --> G
    F --> H[Repaired JSON Stream]

    G --> I[Fact Extractor]
    H --> I

    I --> J[Unified Parsed Fact Records]

Referenced by