What a retrieval corpus gains from one object per line
JSONL is one JSON object per line when vector stores do not share a format. The embedding model belongs in a sibling manifest. RFC 8259 is only the JSON grammar.
The short version
Because there is no shared binary format among vector stores, and one JSON object per line can be streamed, resumed, and inspected without parsing a whole array. The note says the embedding model belongs in a sibling manifest, not in a header line.
What the note is claiming
RAG Converter's note on JSONL says the working shape of a retrieval corpus is one JSON object per line: an identifier, the passage, an embedding as an array of numbers, and metadata for the source file, the page, and the chunk kind. It says Qdrant snapshots, Chroma's SQLite and Parquet, Lance files, and a FAISS index plus a sidecar do not read each other. What they accept, directly or through a small script, is newline-delimited JSON. A single JSON array has to be parsed before the first element can be read. A line can be streamed, a failed import can resume, and the first record can be inspected without loading a file that runs to gigabytes. The paper filing the shape isFutureweb.
What a line cannot say about itself
The note says a JSONL file has no header, and that adding one breaks naive readers. Which embedding model produced the vectors, and whether anything was truncated, belong in a sibling manifest. Without that file, semantic vectors and keyword vectors look the same. The desk isSoftware.
What the standards are, and are not
jsonlines.org documents the convention of one JSON value per line. RFC 8259 is the JSON grammar. Neither document is a vector-database specification, and neither chooses an embedding model. The note is the source for the retrieval shape. The standards are the source for why a line is legal JSON and why a file of lines is a convention rather than an internet standard. What a version number has to mean, for the model named in that manifest, iswhat a version number has to mean.
Why it matters on this desk
The dependency is interchange. A corpus that can move between stores is a software fact. A corpus that cannot name its model is a silent conversion. Chunk size, a different retrieval decision already filed, iswhat the chunk note says you decide before you retrieve.
What this page will not import
This page does not paste the note's product list, and it does not treat JSONL as a formal standard. The mechanism note is the JSONL page. A license obligation, a different open-source fact, iswhat a license obligates.
Where the public record sits
The note, jsonlines.org, and RFC 8259 overlap on one value per record and on JSON as the syntax. Only the note describes embeddings, page numbers, and a sibling manifest. A reader who cites the RFC as if it specified a retrieval corpus is citing the wrong document. A five-line import script is the note's phrase for how a store that lacks a native loader still accepts the lines. This page does not print the script. A five-line import script is the note's phrase for how a store that lacks a native loader still accepts the lines. This page does not print the script. The note says conversion quality in that pipeline comes from pure-Rust parsers. This page does not inventory those parsers. Gigabyte files are the note's reason the streaming property matters. They are not a benchmark. A five-line import script is the note's phrase for how a store that lacks a native loader still accepts the lines. This page does not print the script. The note says conversion quality in that pipeline comes from pure-Rust parsers. This page does not inventory those parsers. Gigabyte files are the note's reason the streaming property matters. They are not a benchmark.
What to watch next
A store that refuses newline-delimited JSON, or a convention that puts the model name in-band without breaking readers, would change the record. Another exporter that emits the same four fields would not. The September 29 argument that a dependency graph is the product is a different software note,what a yellow badge does not review.
What would change the record
A formal standard that vector stores actually implement would change the claim that JSONL won by default. A header line that every reader accepted would change the manifest sentence. A binary snapshot that only one product reads would confirm it. The note says conversion quality in that pipeline comes from pure-Rust parsers. This page does not inventory those parsers. Gigabyte files are the note's reason the streaming property matters. They are not a benchmark. The note says conversion quality in that pipeline comes from pure-Rust parsers. This page does not inventory those parsers. Gigabyte files are the note's reason the streaming property matters. They are not a benchmark. A five-line import script is the note's phrase for how a store that lacks a native loader still accepts the lines. This page does not print the script. The note says conversion quality in that pipeline comes from pure-Rust parsers. This page does not inventory those parsers. Gigabyte files are the note's reason the streaming property matters. They are not a benchmark. A five-line import script is the note's phrase for how a store that lacks a native loader still accepts the lines. This page does not print the script. The note says conversion quality in that pipeline comes from pure-Rust parsers. This page does not inventory those parsers.
What stays a line and a manifest
What stays specific is one JSON object per line, four fields the note says are enough to import and to cite, a sibling manifest for the model, and two public documents that define JSON and the line convention without defining a retrieval corpus. The house link for the mechanism isthe JSONL note.
Sources
The reports this brief is filing. Futureweb did not republish them.
Questions
Why does the JSONL note say a retrieval corpus is line-delimited?
A JSON array must be parsed before the first element. A line can stream, resume after a failed import, and be inspected without loading a multi-gigabyte file.
What four fields does that note say the working record carries?
An identifier, the passage text, the embedding as an array of numbers, and metadata for the source file, the page, and the chunk kind.
Where does the note say the embedding model name belongs?
In a sibling manifest, not in a header line. It says a header breaks naive readers, and that without the manifest semantic vectors and keyword vectors look alike.
Does RFC 8259 specify a vector-store corpus?
No. It specifies JSON. jsonlines.org documents one JSON value per line. The retrieval shape is the note's account, not the RFC's.
Which stores does the note say cannot read each other's files?
It names Qdrant snapshots, Chroma's SQLite and Parquet, Lance files, and a FAISS index plus a sidecar. It says they converge on newline-delimited JSON instead.