perf: cache document structures to avoid repeated disk reads

[mrjunos]

Summary

Each chat request was reading and parsing the PageIndex JSON structure for every indexed document on every call — O(n) disk I/O per request.

Fix: get_structure() now caches results in _structure_cache on first call. Cache is invalidated when a document is re-indexed.

def get_structure(self, doc_id: str) -> str:
    if doc_id not in self._structure_cache:
        self._structure_cache[doc_id] = self.client.get_document_structure(doc_id)
    return self._structure_cache[doc_id]

Tests added (46 total, up from 38)

New test_indexer.py (8 tests):

• First call hits client, second call uses cache
• Cache returns consistent values
• Cache invalidated after re-index
• Independent caching per document
• Deduplication: existing doc not re-indexed, new doc is indexed

Test plan

• 46 tests passing
• CI green

🤖 Generated with Claude Code