-
[OPIK-4493] [FE] Add LangChain/LangGraph message format support (#5301)
frostbyte_neo 发布于
2026-02-24 11:04:52 +00:00 - [OPIK-4493] [FE] Add LangChain/LangGraph message format support
Rename LLMProvider → LLMMessageFormat across the detection system and add
a new "langchain" format so the Messages tab renders for LangGraph traces.Co-Authored-By: Claude Opus 4.6 noreply@anthropic.com
- [OPIK-4493] [FE] Encapsulate combiner logic in llmMessages module
Move detect → map → combine pipeline into mapAndCombineMessages() so
MessagesTab no longer reaches into provider internals. Add FormatCombiner
type and LangChain-specific combiner that deduplicates LangGraph state
messages.Co-Authored-By: Claude Opus 4.6 noreply@anthropic.com
- [OPIK-4493] [FE] Handle nested input wrapper in LangChain detector/mapper
When LangGraph input is a Pydantic BaseModel, the SDK serializes trace
input as { input: { messages: [...] } }. Add unwrapping logic in the
detector and mapper to check data.input.messages when data.messages
is not found (input-only).Co-Authored-By: Claude Opus 4.6 noreply@anthropic.com
- [OPIK-4493] [FE] Fix combiner dropping input when output is not a superset
The LangChain combiner incorrectly treated any flat output.messages as a
LangGraph state superset, dropping input messages. Add a length check so
input is only omitted when output has at least as many messages.Co-Authored-By: Claude Opus 4.6 noreply@anthropic.com
- [OPIK-4493] [FE] Use content fingerprints for combiner dedup instead of raw parsing
Replace raw-message re-parsing in the combiner with a contentFingerprint
computed once during mapping. This fixes false superset detection (equal-length
conversations with different content) and avoids double-parsing of nested input
wrappers.Co-Authored-By: Claude Opus 4.6 noreply@anthropic.com
Co-authored-by: Claude Opus 4.6 noreply@anthropic.com
下载附件