Feature Request: Semantic Search for Traces

[ikeda-tomoya-swx]

Feature Description

Add semantic/vector search capability for traces, allowing users to search traces using natural language queries like:

• "Show me HR-related traces"
• "Find traces about budget questions"
• "Search for error handling conversations"

Current Limitation

Currently, Langfuse supports:

• Full-text search (keyword matching) ✅
• Metadata/tag filtering ✅
• Semantic search ❌

Full-text search requires exact keyword matches, which doesn't work well when users want to find conceptually similar traces without knowing the exact terms used.

Proposed Solution

1. Index trace content in a vector database (e.g., pgvector extension for PostgreSQL, or external like Qdrant/Pinecone)
2. Generate embeddings for trace inputs/outputs during ingestion
3. Hybrid search combining vector similarity with existing filters (time range, tags, user, etc.)

Use Cases

• Debugging: "Find traces where users asked about login issues" (even if they said "can't sign in", "authentication problem", etc.)
• Analysis: "Show me all customer support related conversations"
• Quality Review: "Find traces with negative sentiment"

Prior Art

• Laminar is implementing this with "indexing each span in a vector DB and performing hybrid search at query time"
• This approach is becoming standard in LLM observability platforms

Additional Context

This would be especially valuable for teams with large trace volumes where manual browsing or keyword search is insufficient.

Would love to hear if this is on the roadmap or if there's interest from the community! 🙏

[jannikmaierhoefer]

Hi @ikeda-tomoya-swx, thanks for adding this!

[nimarb]

very interesting idea! as we are storing the trace / observation input/output/metadata in clickhouse, something like this might come in handy: https://clickhouse.com/docs/knowledgebase/vector-search