GraphBuilder ignores AdaptiveBundler output, causing context limit issues

[mzfr]

Hi, I'm sorry If I misunderstood something but right now everything is being crammed into the model context.

I understand that AdaptiveBundler creates X chunks(where X comes from the context set in config). But the GraphBuilder ignores the bundles and tries to use ALL the cards, which might fail(unless you are using openai's 1M context window model).

By fail I mean, when I tried with 15 files and 25K context window, it sampled 1/54 cards

│ 13:15:10 Added: 1 nodes, 2 edges (orphans 0->0)                                                                                                         │
│ 13:15:10 AuthorizationMap: 4N/3E, 0 orphans                                                                                                             │
│ 13:15:10 Context limit reached — sampling active: using 1/54 cards. Consider restricting input files with --files (whitelist), and/or using a model     │
╭────────────────────────────╮
│ ✓ Graph building complete! │

I believe this(d0e4e7e) commit tried fixing that but this(2a8a667) commit brought back the same behaviour.

Was graph-building accuracy taking a significant hit when everything wasn’t kept in a single context, and is that why the change was overridden?

[muellerberndt]

Was graph-building accuracy taking a significant hit when everything wasn’t kept in a single context, and is that why the change was overridden?

Yes exactly. The expected behavior should be to sample cards if not all of them fit in the context, but warn that that graph quality may degrade.

Mainly, schema design is a single prompt and it should ideally see at least 80% of the code. It would be possible to split this into multiple steps to support smaller context windows but I'm not planning to implement this (happy to look at PRs though).

[mzfr]

Sure I'd be happy to make a PR

I was thinking maybe I can do something like

• Load the bundle.json and map card IDs to bundles
• Sample complete bundles rather than randomly selecting individual cards
• Rotate the bundles across iteration, this should expand the coverage
• And maybe fail-fast if cov drops too low (ex: 1/54)

Please let me know your thoughts on this