AI & Agent Dev Bug Sandbox Back to Radar

MarkdownHeaderTextSplitter Suffers O(N²) Slowdown When Merging Many Paragraphs Under The Same Header

aggregate_lines_to_chunks() repeatedly appends to a growing string, causing quadratic time complexity when many adjacent lines share the same metadata. The issue is observable with large inputs (e.g., 16,000 paragraphs take ~5.8 seconds) and becomes progressively worse.

highConfidence 95%Langchain Affected V1.1.2

Origin Analysis

In aggregate_lines_to_chunks(), the code uses `aggregated_chunks[-1]["content"] += " \n" + line["content"]` inside a loop. Each concatenation creates a new string and copies the entire accumulated content, leading to O(n²) time and memory traffic for n lines with identical metadata.

Install langchain-text-splitters==1.1.2. Run the provided benchmark: create text with N paragraphs separated by blank lines and split using MarkdownHeaderTextSplitter(headers_to_split_on=[("#", "Header 1")], strip_headers=False). Time splitting for N = 2000, 4000, 8000, 16000. Observed times: 0.02s, 0.27s, 1.27s, 5.7s respectively, demonstrating super-linear growth.

Fixing Code Block

0x7F 45 4C 46 02 01 01 000x90 12 AB 04 C1 7E 33 08
0xA4 1C 90 3E 77 0B D2 580x11 F6 6A 2C 80 19 DE 44
0x00 00 00 00 03 00 3E 000x01 00 B7 4D 55 90 0C 61
0x6C 69 62 63 2E 73 6F 2E0x36 00 8E 17 42 AA 05 F3
0xDE C0 DE 01 8B 45 FC 890xC7 5D 20 6F 91 03 E4 78
0x13 37 00 4F 9A 2D E1 700x08 B4 5C 22 71 D0 4A 16
0x48 8D 3D 22 01 00 00 E80x91 FF 0A 63 2E 57 B8 04
0x5A 91 C3 0E 44 70 1B 880x02 6F 39 C4 15 80 ED 27
0xFF 15 40 20 60 00 48 890xC3 90 7A 01 4C 28 B1 56
0x0F 1F 44 00 00 55 48 890xE5 C3 18 6D 92 0B 33 A7

Edge Case Audit

The fix introduces a temporary `content_parts` key in the chunk dictionaries. Ensure no downstream code expects only `content` and `metadata` keys in those intermediate dicts. If the method is ever overridden or externally referenced, behavior remains compatible. For rollback, replace this method with the original implementation, but re-run performance tests to confirm old behavior. No concurrency or multi-threading concerns are introduced. Test with empty lines, single lines, and mixed metadata to verify identical output.