Skip to content

Mistral OCR to LanceDB: Chunking for RAG ​

The short answer: Mistral OCR gives you excellent per-page markdown; LanceDB gives you an embedded, Arrow-native vector table with no server to operate. Wired together naively — flatten, split, embed, insert — they still produce mediocre RAG, and a naive row shape can also break LanceDB's schema inference across documents. The missing link is POMA: PrimeCut().ingest() consumes the raw /v1/ocr JSON, emits hierarchy-preserving chunksets, and vektoria's content-free ingest pattern writes a fixed, small row shape into LanceDB while assemble() reconstructs prompt-ready context from a volume at query time.

What each end of the pipeline actually provides ​

Mistral OCR (/v1/ocr, mistral-ocr-latest) returns pages[], each with an index and a markdown string — headings, lists, and tables recovered within each page, image bytes inline when you set include_image_base64, in an images array whose length varies page to page. What it doesn't provide: cross-page hierarchy (a heading on page 41 has no machine link to the chapter that opened on page 3) or retrieval units. Details: The Optimal Chunker for Mistral OCR.

LanceDB provides an embedded, Arrow/Lance-backed table — local or s3://-backed, no server — with a vector column, SQL-like .where(...) filtering on scalar columns, and an optional full-text index for hybrid queries. What it doesn't provide: any opinion about what a row should hold, and its table schema is inferred once, from the first batch written. Details: LanceDB Chunking Strategy for RAG.

The naive wiring, and where it breaks ​

The common recipe — "\n".join(p["markdown"] for p in pages) → RecursiveCharacterTextSplitter → embed → build a row dict straight from each page's raw fields → create_table(...) — breaks in a pair-specific way:

  • Mistral's page indices vanish at the join, so no scalar column can carry a page number, and .where(...) has only file_id left to filter on.
  • Overlap adds near-duplicate rows that inflate both the table and the vector index built over it, for no retrieval benefit.
  • LanceDB's schema is fixed at create_table, inferred from the first batch — and Mistral's own output shape is exactly what breaks it. If a naive pipeline keeps Mistral's raw images field (or any other page-specific key) on the row, the column set differs between a page with two images and a page with none, or between one document and the next. LanceDB either rejects the mismatched add() call or silently coerces the data against a schema that no longer describes it — a failure mode unique to LanceDB's Arrow-first design, and one a content-free row shape (only id, vector, file_id, chunkset_index) sidesteps completely, since those four fields never vary across pages or documents.

The pipeline, end to end ​

bash
pip install mistralai lancedb
python
import os
from mistralai.client import Mistral  # mistralai>=2 is a namespace package; the client lives in mistralai.client
import lancedb
from poma import PrimeCut
from poma.vektoria import assemble, records_from_archive, Volume
from poma.embeddings import get_embedder

# 1. Mistral OCR — your existing call, unchanged. Save the raw result JSON.
mistral = Mistral(api_key=os.environ["MISTRAL_API_KEY"])
ocr = mistral.ocr.process(
    model="mistral-ocr-latest",
    document={"type": "document_url", "document_url": "https://example.com/contract.pdf"},
    include_image_base64=True,
)
with open("contract.mistral-ocr.json", "w") as f:
    f.write(ocr.model_dump_json())

# 2. The missing link — raw result JSON in, a portable .poma archive out.
client = PrimeCut()  # reads POMA_API_KEY
client.ingest("contract.mistral-ocr.json", download_dir="archives", filename="contract.poma")

embedder = get_embedder("local:BAAI/bge-small-en-v1.5")
vol = Volume("s3://your-bucket/poma")
db = lancedb.connect("s3://your-bucket/lancedb")

# 3. Content-free ingest — a fixed row shape, so schema inference never breaks across documents.
records = records_from_archive("archives/contract.poma")  # -> list[Record] (id, text, payload)
rows = []
# assemble() reads the FULL document (chunks + chunksets) from the volume — write it once,
# not one {"chunks": <indices>, "text"} stub per record (which also overwrote itself each turn).
from poma.utils import unpack_poma_archive
_data = unpack_poma_archive(poma_archive_path="archives/contract.poma")
file_id = records[0].payload["file_id"]
vol.write_doc(file_id, {"file_id": file_id, "chunks": _data["chunks"], "chunksets": _data["chunksets"]})
for r in records:
    rows.append({
        "id": r.id,
        "vector": embedder.embed([r.text])[0],
        "file_id": r.payload["file_id"],
        "chunkset_index": r.payload["chunkset_index"],
    })

tbl = db.create_table("poma", data=rows, exist_ok=True)

# 4. Vector search retrieval — metadata returned by default, then assemble.
qv = embedder.embed(["early termination conditions"])[0]
results = tbl.search(qv).where(f"file_id = '{file_id}'").limit(10).to_list()  # POMA assigns an opaque file_id, not the file name
context = assemble(results, volume=vol)  # -> [{"file_id", "content"}, ...]

POMA validates the Mistral payload shape up front (a corrupted or mislabeled upload 422s immediately), handles all three Mistral image cases, strips running headers/footers, and rebuilds the cross-page heading tree before chunking. Retrieved chunksets share ancestor lineage across a document — assemble() deduplicates that lineage into one prompt-ready cheatsheet, the same discipline that answers our reference legal-document benchmark with 337 tokens instead of 1,542 for a recursive-splitter baseline. Methodology here.

Metadata mapping: POMA fields → LanceDB primitives ​

POMA chunk fieldLanceDB primitiveWhat it enables
to_embedvector columnANN search via tbl.search(...)
file_idscalar column.where("file_id = '...'") scoping
page (from Mistral pages[].index)volume document, not a columncontent-free retrieval, cheatsheet assembly
depthvolume document, not a columncontent-free retrieval, cheatsheet assembly
chunkset_indexscalar column, returned by defaultassemble() dedup and ordering
chunkset lineage (content)not stored in the table — lives on the Volumecontent-free table, no schema churn from raw page fields

Frequently asked questions ​

How do I get Mistral OCR results into LanceDB for RAG? ​

Save the raw /v1/ocr JSON, run PrimeCut().ingest() on it, then pull records_from_archive(...), write each chunkset's content to a volume, embed its to_embed text, and create_table/add rows (id, vector, file_id, chunkset_index). Retrieve with tbl.search(...).to_list() and assemble().

Does POMA's assemble() need any special flag to read LanceDB results after a Mistral OCR ingest? ​

No — LanceDB returns every column by default on .to_list(), including file_id and chunkset_index, so assemble(results, volume=VOL) works with no extra request. LanceDB is the one new-DB pair here with no metadata-return gotcha.

What happens if I insert raw Mistral OCR page dicts straight into a LanceDB table? ​

LanceDB infers a table's schema from the first batch written. Mistral's images array varies in length page to page and document to document, so a row shape built from raw page dicts changes columns across inserts, and later add() calls fail or coerce against a stale schema. A fixed, content-free row shape avoids this.

Why not just split Mistral OCR's markdown and add it to a LanceDB table directly? ​

Concatenation discards page indices and heading levels, and splitter overlap adds near-duplicate rows that inflate the table and its vector index for no retrieval benefit. LanceDB then ranks context-free fragments exactly as fast as it would rank real chunksets.

Is LanceDB a good fit for prototyping a Mistral OCR RAG pipeline before scaling it? ​

Yes — its embedded, Arrow/Lance-backed design needs no server, and tables scale to object storage without a schema rewrite, so a pipeline built with the content-free chunkset row shape from day one carries into a larger deployment unchanged.

Same parser, different store: Mistral OCR → Chroma · Mistral OCR → Qdrant · Mistral OCR → pgvector

Foundations: The Optimal Chunker for Mistral OCR · LanceDB Chunking Strategy for RAG · All pipeline recipes · RAG chunking guide · POMA chunksets