doc4d-2026-08-05

A vector-search dataset built from the 4D Documentation website. Each row is a chunk of plain text extracted from a documentation page, paired with a 1024-dimensional embedding, for semantic (meaning-based) search over url + text.

Dataset Details

Dataset Description

This dataset contains text chunks scraped and extracted from the 4D Documentation site (https://developer.4d.com/), embedded with LFM2.5-Embedding-350M (quantized Q8_0 GGUF), and packaged both as a raw JSONL source file and a ready-to-query SQLite vector database using sqlite-vec.

The intended use is semantic retrieval: given a natural-language query, embed it with the same model, run a nearest-neighbor search over the embedding column, and return the matching url and text (with language and version as metadata) so the source documentation can be located and cited.

• Curated by: keisuke-miyako
• Language(s): Text extracted from 4D Documentation pages (multiple locales, e.g. en, fr — see language column for the actual set present)
• Source: Plain text extracted from HTML pages on https://developer.4d.com/
• License: Not specified — inherits whatever license/terms apply to 4D Documentation content itself; check with 4D / 4D SAS before redistribution or commercial use.

Dataset Sources

• Repository: https://huggingface.co/datasets/keisuke-miyako/doc4d-2026-08-05
• Original documentation site: https://developer.4d.com/

Files

File	Description
data.jsonl	Source data. One JSON object per line, each a single documentation text chunk with its embedding.
doc.db	SQLite database with a sqlite-vec virtual table (chunks) built from data.jsonl, ready for nearest-neighbor queries.

data.jsonl schema

Each line is a JSON object:

Field	Type	Description
embedding	list[float], length 1024	Embedding vector for text, generated by LFM2.5-Embedding-350M-Q8_0.
url	string	Source URL of the documentation page the chunk was extracted from.
language	string | null	Language of the page/chunk, if known.
version	string | null	4D product version the page applies to, if known.
text	string	Plain text of the chunk (HTML stripped).

doc.db schema

Built with the sqlite-vec extension as a vec0 virtual table:

CREATE VIRTUAL TABLE chunks USING vec0(
    embedding float[1024] distance_metric=cosine,
    +url TEXT,
    +language TEXT,
    +version TEXT,
    +text TEXT
);

• embedding is indexed for approximate/exact nearest-neighbor search (cosine distance).
• url, language, version, text are stored as auxiliary (+) columns, returned alongside search results but not searched directly.

The database was built from data.jsonl with a straightforward batch-insert script (float32 little-endian packing via Python's struct, sqlite_vec.load() to register the extension, batched commits of 2000 rows at a time).

Embedding Model

• Model: LiquidAI/LFM2.5-Embedding-350M-GGUF
• Quantization: Q8_0 (LFM2.5-Embedding-350M-Q8_0.gguf)
• Embedding dimension: 1024
• Distance metric: cosine

To query this dataset correctly, encode queries with the same model and quantization, since embeddings from a different model or checkpoint will not be comparable in the same vector space.

Uses

Direct Use

• Semantic search over 4D Documentation: embed a natural-language question, retrieve the most similar chunks, and use the returned url/text to answer questions or point users to the right documentation page.
• Building a retrieval-augmented generation (RAG) pipeline for 4D-related developer support or chat assistants.

Out-of-Scope Use

• Not intended as a general-purpose or up-to-date mirror of 4D Documentation — content reflects a snapshot as of the dataset's build date (2026-08-05) and may be stale.
• Not suitable for embeddings-space comparison against other embedding models without re-encoding.

How to Use

Query with sqlite-vec (Python)

import sqlite3
import struct
import sqlite_vec

DIM = 1024

def serialize(vec):
    return struct.pack(f"{len(vec)}f", *vec)

db = sqlite3.connect("doc.db")
db.enable_load_extension(True)
sqlite_vec.load(db)
db.enable_load_extension(False)

# query_embedding: list[float] of length 1024, produced by the
# same LFM2.5-Embedding-350M-Q8_0 model used to build this dataset
q_blob = serialize(query_embedding)

rows = db.execute(
    """
    SELECT url, text, language, version, distance
    FROM chunks
    WHERE embedding MATCH ?
    ORDER BY distance
    LIMIT ?
    """,
    (q_blob, 10),
).fetchall()

for url, text, language, version, distance in rows:
    print(distance, url)

Load the raw source with 🤗 Datasets

from datasets import load_dataset

ds = load_dataset("keisuke-miyako/doc4d-2026-08-05", data_files="data.jsonl", split="train")
print(ds[0]["url"], ds[0]["text"][:200])

Note: the embedding field will load as a plain list of floats via this path; for actual vector search, use doc.db directly with sqlite-vec as shown above rather than a brute-force scan over data.jsonl.

Dataset Creation

Source Data

Text was extracted from HTML pages published on the 4D Documentation site and split into chunks. Each chunk retains its source url, and, where available, language and 4D product version metadata.

Data Collection and Processing

1. HTML pages from developer.4d.com were fetched and converted to plain text.
2. Text was chunked and each chunk embedded with LFM2.5-Embedding-350M-Q8_0 (1024-dim).
3. Chunks + embeddings + metadata were written to data.jsonl.
4. data.jsonl was loaded into a sqlite-vec vec0 virtual table (chunks) in doc.db, batched in groups of 2000 rows, using cosine distance as the similarity metric.

Who Produced This Data

Compiled by keisuke-miyako from publicly available 4D Documentation content.

Bias, Risks, and Limitations

• Snapshot, not live data: the dataset reflects the state of 4D Documentation at build time (2026-08-05) and will drift out of date as the docs are updated.
• Coverage: only pages that were crawled/extracted are represented; gaps in language or version metadata may exist for some chunks (null values).
• Attribution: documentation text is owned by 4D SAS; this dataset redistributes extracted text for search/retrieval purposes. Confirm usage rights before commercial or redistribution use.
• Embedding staleness: if LFM2.5-Embedding-350M is updated or superseded, embeddings here will not match newer versions of the model.

Citation

If you use this dataset, please cite the repository:

@misc{doc4d2026,
  title = {doc4d-2026-08-05: Semantic search dataset for 4D Documentation},
  author = {keisuke-miyako},
  year = {2026},
  howpublished = {\url{https://huggingface.co/datasets/keisuke-miyako/doc4d-2026-08-05}}
}

Also consider citing the embedding model:

@misc{lfm25embedding,
  title = {LFM2.5-Embedding-350M},
  author = {LiquidAI},
  howpublished = {\url{https://huggingface.co/LiquidAI/LFM2.5-Embedding-350M-GGUF}}
}

Dataset Card Contact

For questions about this dataset, open a discussion on the repository.