fix(kb): clamp markdown chunk overlap to the heading-reduced body budget

[iuiu-py]

Fixes #9998

What Changed

In MarkdownChunker._sections_to_chunks(), when a section's heading-path prefix reduces the body's effective chunk_size, the configured chunk_overlap is now clamped to that reduced budget (max(0, min(chunk_overlap, effective_chunk_size - 1))), and the effective budget itself is floored at 1. Sections without heading context and sections that fit in one chunk are unaffected.

Why

The chunker deducts the heading-prefix length from the body budget but passed the caller's chunk_overlap through untouched. The overlap is validated externally against the original chunk_size, so an externally legal pair — e.g. chunk_size=256, chunk_overlap=100 with a 200-char parent heading — became overlap(100) >= effective_chunk_size(64) inside the recursive chunker, which raises ValueError: chunk_overlap must be less than chunk_size and fails the whole knowledge-base upload at the chunking stage. The exact same input chunks fine with include_heading_context=False, which made the failure look config-independent.

Testing

• tests/unit/test_markdown_chunker_overlap.py::test_long_heading_prefix_clamps_overlap_to_body_budget reproduces the issue verbatim ("# " + "A"*200 + "\n\n## Child\n" + "B"*600, 256/100): it raises on unpatched master (95e98b8) and produces body-bearing chunks with the fix.
• The control case (include_heading_context=False) is pinned by the second test.
• python -m pytest tests/unit/test_markdown_chunker_overlap.py -q → 2 passed; ruff format/check clean on both files.

Summary by Sourcery

Fix knowledge-base chunking and retrieval ranking edge cases, and produce valid streamed MiniMax WAV output.

New Features:

• Repair streamed MiniMax WAV audio headers so generated files remain compatible with strict decoders.

Bug Fixes:

• Prevent markdown chunking failures when heading context reduces the effective body budget below the configured overlap.
• Ensure candidates returned by only one retrieval channel are not unfairly penalized during rank fusion.

Tests:

• Add coverage for markdown overlap clamping, retrieval-channel-only ranking, and MiniMax WAV header repair.

[sourcery-ai]

Hey - I've found 1 issue

Prompt for AI Agents

Please address the comments from this code review:

## Individual Comments

### Comment 1
<location path="astrbot/core/provider/sources/minimax_tts_api_source.py" line_range="42-43" />
<code_context>
+        if csize > remaining:
+            break
+        pos = chunk_start + csize + (csize % 2)
+    if not fmt or not data_start or not data_size:
+        return audio
+    if (
+        struct.unpack_from("<I", audio, 4)[0] == len(audio) - 8
</code_context>
<issue_to_address>
**issue (bug_risk):** A valid zero-length WAV, or a truncated streamed WAV whose data chunk has no remaining payload, takes the early return and retains a stale `0xFFFFFFFF` data length instead of being repaired. `get_audio` accepts any nonempty returned bytes and writes that invalid header to disk.

**Triggers:** When the assembled stream has a `fmt ` chunk followed by a zero-byte `data` payload and an inconsistent data-size field.

**Suggested fix:** Distinguish absence of a `data` chunk from a present zero-length `data` chunk, and repair the latter even when its actual payload length is zero.
</issue_to_address>

Sourcery assessment

Needs a human reviewer. 1 finding to address first, and the markdown change can produce different knowledge-base chunks and embeddings that remain stored after a revert, so correcting a wrong result may require reindexing, although the impact is bounded and repairable. The rank-fusion and WAV-header changes affect runtime retrieval or playback behavior and are otherwise reversible by reverting the PR.

Blocking findings: astrbot/core/provider/sources/minimax_tts_api_source.py:43

---

Sourcery is free for open source - if you like our reviews please consider sharing them ✨

• X
• Mastodon
• LinkedIn
• Facebook

[sourcery-ai]

issue (bug_risk): A valid zero-length WAV, or a truncated streamed WAV whose data chunk has no remaining payload, takes the early return and retains a stale 0xFFFFFFFF data length instead of being repaired. get_audio accepts any nonempty returned bytes and writes that invalid header to disk.

Triggers: When the assembled stream has a fmt chunk followed by a zero-byte data payload and an inconsistent data-size field.

Suggested fix: Distinguish absence of a data chunk from a present zero-length data chunk, and repair the latter even when its actual payload length is zero.