Incremental semantic cache checkpoint fails on sliced docs: _checkpoint_chunk reads FileSlice.rel, which does not exist

[mphstudios]

Summary

During graphify extract, the per-chunk incremental semantic cache checkpoint fails for every chunk that contains an oversized (sliced) document, printing:

[graphify] incremental cache checkpoint failed: argument should be a str or an os.PathLike object where __fspath__ returns a str, not 'FileSlice'

Extraction completes and produces a valid graph — but the incremental cache the checkpoint exists to provide is silently defeated for those chunks, so a re-run (or a run resumed after a crash/kill/rate-limit) re-bills them instead of reusing cached results.

Environment

• graphify (graphifyy) 0.9.14, Python 3.13, macOS.
• Observed on graphify extract . --backend claude-cli over a 70-document corpus (5 chunks): chunks 1–4 each printed the failure; chunk 5 (no sliced doc) did not.

Root cause

llm.py, _checkpoint_chunk builds the allowed_source_files list as:

# A FileSlice reports its file via `.rel`; a bare Path is the relative source_file itself.
allowed = [getattr(item, "rel", None) or item for item in chunk]

But FileSlice (file_slice.py) is a @dataclass(frozen=True) whose path field is .path — it has no .rel attribute ([a for a in dir(FileSlice) if not a.startswith('__')] is empty of rel). So for every FileSlice, getattr(item, "rel", None) returns None, and or item leaks the FileSlice object itself into allowed.

save_semantic_cache(..., allowed_source_files=allowed) then does, in resolved_source_path:

path = Path(value)   # value is a FileSlice -> TypeError

(cache.py), which raises argument should be a str or an os.PathLike object ... not 'FileSlice'. The except Exception in _checkpoint_chunk catches it and prints the best-effort warning, so extraction continues but the chunk is never checkpointed.

The comment at the call site referencing .rel is stale relative to the FileSlice dataclass.

Reproducer

graphify extract . --backend <any> --out <dir> on a corpus containing at least one document large enough to be split into FileSlice units (#1369). Every chunk that includes a slice prints the checkpoint failure.

Suggested fix

Resolve a FileSlice's source path via its actual attribute (.path) rather than the non-existent .rel — e.g. getattr(item, "path", None) or item, or unwrap FileSlice explicitly before building allowed. Whatever attribute genuinely carries the relative source_file should be used; .rel does not exist on the current dataclass.

[HerenderKumar]

Working on this — PR shortly.

[HerenderKumar]

PR up: #1882.

[safishamsi]

Fixed in cfc7cf2 (ships in 0.9.16). _checkpoint_chunk now resolves each unit via unit_path (was reading the non-existent FileSlice.rel), so sliced-document chunks checkpoint correctly.