Better Vector Search for Long Documents: Chunking Inside Manticore Search
Manticore Search added automatic document chunking for vector columns, lifting long-document recall@5 from 55.1% to 83.3% in its benchmarks.
Manticore Search introduced a chunk_strategy option for model-backed vector columns in CREATE TABLE, offering five strategies (truncate, mean, fixed, recursive, sentence) with tunable max_tokens, overlap_tokens, and max_chunks, eliminating external splitters and separate chunk tables. On its 189-page, ~298k-word manual, sentence chunking improved recall@5 from 55.1% to 83.3% and MRR from 0.44 to 0.70, at roughly 2.5x RAM and 4x ingest time. Documents still return as single results; queries are never chunked.
- chunk_strategy is set directly in CREATE TABLE on model-backed vector columns
- fixed, recursive, and sentence strategies require float_vector_array columns
- knn_dist() reports distance to the closest chunk; k counts documents
- Trade-off measured at ~2.5x RAM and ~4x ingest time
- Vendors
- Manticore Search
- Products
- Manticore Search
- AI models
- Xenova/all-MiniLM-L6-v2
From the original article
14 min read at manticoresearch.com
Say you are building search over your team's internal documentation — guides, runbooks, postmortems. You have a table with auto embeddings : you insert text, Manticore runs the model and fills the vector column for you. (If that is new to you, start with vector search in Manticore .)
By GloriaVinogrado · manticoresearch.com
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.