{
  "id": 8009371,
  "title": "Better Vector Search for Long Documents: Chunking Inside Manticore Search",
  "url": "https://urgent.news/2026/09/17/better-vector-search-for-long-documents-chunking-inside-manticore",
  "topic": "tech",
  "section": "Tech",
  "published": "2026-09-17T10:30:52.000Z",
  "source": {
    "name": "Hacker News",
    "slug": "hacker-news",
    "url": "https://manticoresearch.com/blog/auto-chunking/"
  },
  "original_language": "en",
  "account": "Manticore Search introduces chunking for long documents, allowing for more accurate vector search results. Previously, if a document exceeded the model's 512-token input window, it would only consider the first 380 words and discard the remaining content, resulting in incomplete search results. With the new chunking feature, the table definition can include chunk_strategy, which splits each document into chunks, embeds every chunk, and searches all of them, eliminating the need for manual chunking or additional storage tables.\n\nTo use chunking, create a table with three vector columns, each representing a different chunking strategy. Insert documents using a single INSERT statement, which fills all three vector columns. When querying a question answered within the document, the runbook's last section containing the certificate-rotation paragraph yields a higher relevance score when using sentence chunking compared to the truncate strategy, which only considers the entire document as a whole.\n\nChunking offers several benefits over the default truncate strategy. It captures relevant information more accurately by storing nine vectors instead of one, with each vector representing a chunk containing a specific paragraph. This approach ensures that each chunk competes independently, and Manticore returns the document once, scored by its best chunk. Additionally, chunking is suitable for long-form content such as documentation pages, contracts, transcripts, and email threads where a single document may cover multiple topics.\n\nHowever, chunking should not be used for documents with unrelated topics, as averaging the chunk vectors may produce a representation that is close to neither topic. In such cases, using truncate strategy is recommended, as it performs better in terms of storage and search costs and minimizes the risk of diluting relevant information.\n\nWhen using chunking, the max_tokens parameter determines the maximum number of tokens per chunk. The index size is predictable, as the chunk count depends on the document length. Overlap_tokens can be adjusted to mitigate the issue of a boundary landing mid-sentence, which may affect the embedding quality. The sentence chunking strategy is designed for human-written content with clear paragraph structure, such as Markdown, HTML, wiki pages, and knowledge bases. It detects sentence boundaries using Unicode UAX #29 and greedily packs whole sentences until the token budget is reached. This strategy is particularly useful when the chunks will be fed to an LLM later, as it ensures that each chunk reads coherently and avoids mid-clause splits.",
  "summary": null,
  "key_points": [],
  "editors_take": null,
  "illustration": null,
  "coverage": {
    "outlets": 1,
    "also_reported_by": []
  },
  "ai_generated": true,
  "disclaimer": "Summaries, key points and the editor’s take are written by software from other outlets’ reporting and may contain errors — always check the linked original."
}