Attempted to run backlog ingestion (1,973 pending messages) using NUC qwen3-1.7b pipeline. Hit context size limit errors immediately against production graph.
- Sandbox benchmark (Mar 6): 10/10 success with qwen3-1.7b at 32K context
- Production graph size: 19,841 messages already ingested
- Error:
Error code: 400 - {'error': 'Context size has been exceeded.'}
The 1.7b model worked perfectly during benchmarking because it was tested against a small sandbox graph. Against production with accumulated graph context (nodes, edges, relationship history), the 32K context window is insufficient for Graphiti's retrieval + extraction workflow.
Check if LM Studio config for qwen3-1.7b can be raised beyond 32K. Model architecture may support larger context.
- Available on NUC
- Benchmark showed 75% JSON validity (vs 100% for 1.7b)
- Would need validation run before production use
- Tradeoff: reliability vs capacity
Reduce accumulated context by:
- Deduplicating edges (known issue per TODO.md)
- Pruning low-value nodes
- Clearing entity overlap noise This might bring context back within 32K range.
- Wait for OpenAI quota reset
- Known reliable, but costs $ and has quota limits
- Doesn't solve root issue
This requires Jeff's input on tradeoffs. The graph curation (option 3) was already flagged in TODO.md as needed work. May be the right first step before resuming ingestion.
Ingestion halted. 1,973 messages remain pending. No data corruption - safe to resume once approach is chosen.