Tracing in Ahnlich AI

Why would you want to trace?

With AI tracing, you can:

• Measure embedding computation times for different models (e.g., resnet-50 vs all-MiniLM).

• Track full similarity searches (GETSIMN) including preprocessing and DB lookups.

• Debug performance issues with execution providers or index algorithms.

• Correlate AI spans with DB spans for full pipeline visibility.

How we support distributed tracing

• Enable tracing: --enable-tracing.

• Forward spans: --otel-endpoint http://jaeger:4317.

• Each AI action generates spans for:

  • Preprocessing

  • Embedding generation

  • Similarity algorithms

  • Downstream DB calls

Example docker-compose.yaml

ahnlich_ai:
  image: ghcr.io/deven96/ahnlich-ai:latest
  command: >
    "ahnlich-ai run --db-host ahnlich_db --host 0.0.0.0 \
    --supported-models all-minilm-l6-v2,resnet-50 \
    --enable-tracing \
    --otel-endpoint http://jaeger:4317"
  ports:
    - "1370:1370"

Viewing AI Traces in Jaeger

1. Open http://localhost:16686.

2. Select service = ahnlich_ai.

3. Click Find Traces.

4. Example trace breakdown for GETSIMN:

   • Span 1: Request received by AI.

   • Span 2: Preprocessing (RawString → embedding input).

   • Span 3: Embedding model execution (resnet-50).

   • Span 4: Similarity algorithm execution (cosinesimilarity).

   • Span 5: Predicate filtering in Ahnlich DB (cross-service span).

This gives you a timeline view of the query, where each bar represents time spent in different stages.