Summary

• When index_resource calls vectorize_file without a summary in summary_dict, the abstract field on the Context is set to an empty string
• This means L2 (leaf) records in the vector database have empty abstract fields
• Downstream, hierarchical_retriever passes these empty abstracts as documents to the rerank API, causing rerank providers (e.g. DashScope qwen3-rerank) to return HTTP 400 because they reject empty document strings

Root cause

index_resource (line 371) calls vectorize_file(summary_dict={"name": file_name}) — no "summary" key. Inside vectorize_file, summary = summary_dict.get("summary", "") resolves to "", which becomes Context(abstract=""). The file content IS read for embedding but never used to populate abstract.

Fix

When vectorize_file reads raw file content for embedding and the abstract is still empty, backfill it with the first 200 characters of the file content:

content = _truncate_text(content)
if not context.abstract and content:
    context.abstract = content[:200]
context.set_vectorize(Vectorize(text=content))

Impact

Every L2 record created via index_resource will now have a non-empty abstract, preventing rerank 400 errors.

Test plan

• Run index_resource on a directory with text files
• Verify L2 records in vectordb have non-empty abstract
• Run a search query that triggers rerank — confirm no 400 errors

🤖 Generated with Claude Code