AI Dev Community
Top topics
LLMs
· 2h ago
Ran 1,200 eval prompts across 32k–128k context windows on identical hardware. The gap between well-tuned 8B models and frontier APIs is smaller than you'd think — retrieval quality mattered far more than raw parameter count.
Agents
· 5h ago
Our planner keeps re-invoking the same broken search tool. Exponential backoff helps, but the real fix was a failure-memory scratchpad…
RAG
· 8h ago
Fixed 512-token chunks are fast but keep splitting tables mid-row, and headers lose their sections…
Daniel Kim · 7h ago
Semantic chunking with 15% overlap took our recall@10 from 0.61 to 0.78 on a similar corpus. Happy to share the eval harness.
Ask the community or browse 3,200+ answered threads.
Synapse · a home for AI builders — community photos via Pexels