Lossless JSON Parsing (parseJsonPrioritised)#

parseJsonPrioritised() is a precision-safe JSON parser in the shared utilities package (packages/shared/src/utils/json.ts). It solves a specific problem: JavaScript's native JSON.parse silently rounds integers outside the IEEE 754 safe-integer range (±2⁵³), which corrupts large numeric IDs and timestamps common in LLM traces, dataset items, and evaluation metadata.

How It Works#

The function uses a two-path strategy 1:

1. Fast path — if the input string contains no sequence of 13+ digit/dot characters and no scientific notation (regex UNSAFE_NUMBER_PATTERN = /[\d.]{13,}|\d[eE]/), it delegates directly to native JSON.parse. This keeps the common case cheap.

2. Slow path — if potentially unsafe numbers are detected, it invokes parse() from lossless-json with a custom number reviver:

   • Numbers passing isSafeNumber() are converted to JS Number (normal behavior).
   • Numbers failing isSafeNumber() are preserved as their string representation (e.g., "107505301260286111"), preventing silent rounding.

Parse errors are caught and the original string is returned as a fallback 2.

Where It's Used#

The function is exported from the shared package index 3 and consumed by backend workers, the API server, and the web frontend.

Key Design Decisions#

• Strings, not BigInt — unsafe integers are kept as strings rather than bigint because the rest of the pipeline (ClickHouse, LLM prompts, Lambda/runtime boundaries, JSON serialization) is JSON-based. bigint would not survive re-serialization 4.
• Regex pre-screening — the UNSAFE_NUMBER_PATTERN check avoids invoking the slower lossless-json parser on the majority of payloads that contain only safe numbers 5.
• Context: eval precision fix — PR #14119 extended parseJsonPrioritised usage to the full eval/dataset pipeline after discovering that JSON.parse was corrupting large dataset integer IDs (e.g., 107505301260286111) before they reached LLM prompts.

Related Utilities in the Same File#

• deepParseJson / deepParseJsonIterative — recursively unwrap nested stringified JSON up to configurable depth and size limits. These use native JSON.parse internally (not lossless), so they are suited for UI rendering rather than precision-sensitive ingestion.
• tryParsePythonDict — converts Python dict/list syntax (e.g., LangChain/LangGraph tool calls) to valid JSON before parsing.