Context
#2
by
wolframko - opened
How does it work with codebase exploration? Its maximum context is only 16k tokens. When I use models like Codex or Sonnet, they often consume hundreds of thousands of context tokens. How does this model manage?
@wolframko LocoOperator-4B acts as a sub-agent (like Haiku) to handle the legwork for the main agent (Sonnet). It is not meant to replace the main agent.
In practice, we deploy it with 50K context (30K-50K is the sweet spot), which is more than enough for ~10 turns of targeted codebase searching. It doesn't need to hold the whole codebase; it just finds what Sonnet needs.
Check the architecture details in the README to see how the two-tier system works: https://github.com/LocoreMind/LocoOperator