mdcx

DOI

Convert a document collection to verified Markdown, package it into a single encrypted archive with its search index and provenance, and serve it to agents over the Model Context Protocol.

Contents

• Scope
• Requirements
• Installation
• Quick start
• Conversion
• Packaging
• Querying
• Retrieval model
• Calibration
• Vocabulary
• Transcription recovery
• Expressions
• Quotations across a cut
• Attachments
• Dates
• Correspondence
• Incremental packaging
• Resident conversion
• MCP server
• Sources
• Portable paths
• Signing
• Encryption
• Package format
• Limitations
• Tests
• Contributing
• Security
• Releases
• Authorship
• Citation
• Licence

Scope

mdcx addresses one constraint. An agent asked a question about a document collection must either receive the documents in its context window, which is bounded and billed per token, or query a component that holds an index and returns only the passages that bear on the question. mdcx implements the second.

Three properties define the result:

• Fidelity is checked. Each conversion is compared against the text the original exposes and the coverage achieved is recorded per file. Files that expose no text are marked unverifiable rather than reported as complete. What the comparison establishes differs by engine, and Verification states which.
• The corpus is one artefact. Passages, index and provenance are held in a single AES-256-GCM file whose header can be read without the key, and which may be signed.
• Answers are citable. Every passage carries its source document and its position, so an answer can be quoted against a location.

The pipeline has three stages. Conversion attempts each document with the least expensive engine able to read it and escalates only where that engine falls short: direct text extraction, then a pass that recovers drawn tables, then layout analysis, and optical character recognition for documents that expose no text. Packaging writes the corpus, its index and the provenance of each passage to one file. Retrieval merges word matching and dense retrieval by reciprocal rank, and states when nothing in the corpus bears on the question rather than returning its nearest passage.

Requirements

Python 3.11 or later. No other component is required to query a package.

The floor is 3.11 because a package is one SQLite database, loaded from the decrypted bytes when it is opened, and sqlite3 gained the required call in that version. Conversion and cross-language retrieval each add dependencies, listed under Installation.

Installation

Querying and conversion are separated because their requirements differ by roughly two orders of magnitude. A recipient who only reads packages installs neither Docling nor PyTorch.

The multilingual extra supplies the embedding model, downloaded once on first use, and is required only for queries that cross languages.

The tables extra covers tables that are not drawn with rules — a screenshot of a spreadsheet, or a layout held together by alignment. It reads the geometry only; cell contents still come from the text layer of the document. Without it those pages are handled by the layout engine.

Systems with a CUDA device

Install mdcx[all-gpu] rather than mdcx[all], and install onnxruntime-gpu separately.

onnxruntime and onnxruntime-gpu publish the same module and cannot coexist: whichever was installed last takes effect. An extra that pins the CPU build will therefore displace an accelerated one on every upgrade. all-gpu is all without that pin.

pip cannot express a dependency satisfied by either distribution, so this cannot be settled by declaration alone. mdcx-convert checks at startup and reports when the machine has a device the runtime does not offer. The check on its own:

python -c "import onnxruntime as o; assert 'CUDAExecutionProvider' in o.get_available_providers()"

Quick start

pip install "mdcx[convert]"

mdcx-convert --input ./Documents --output ./Documents_md
mdcx pack --output ./Documents_md --target corpus.mdcx --key "passphrase"
mdcx search corpus.mdcx "where is the storage temperature stated" --key "passphrase"

Conversion

mdcx-convert --input ./Documents --output ./Documents_md

The output folder mirrors the input directory structure, one .md per source document, together with an index of the run.

An interrupted run resumes: work already recorded is not repeated unless --force is given.

Long documents are split into chapters, each converted independently, with an index document recording the correspondence to the original. A chapter is converted from an extract of the document, and its page markers state the document's own numbering rather than the extract's; its front matter declares first_page and last_page, so a passage can be cited back to the page it came from. The index document declares type: document_index and holds no text of its own — the text is in the chapters. --sample-pages converts a spread sample instead of the whole document; the front matter records both the pages taken and the document's total, so a sample is not mistaken for a short document.

--timeout exists because some documents do not complete under layout analysis. Without a limit one such document holds a worker for the length of the run. --recycle-after replaces a worker periodically, since abandoning a document does not stop the thread it started.

Formulas

The text layer of a PDF does not carry the structure of a formula: the bar of a fraction is a drawn stroke, and a superscript is loose text on another line. A faithful extraction of that layer therefore renders a quotient as a single line. Nothing has gone wrong — the structure was never there to extract.

--formulas transcribes the formulas to LaTeX, which means reading their image rather than their text. That is recognition rather than extraction: a different kind of claim about the document, and one that can be wrong in ways extraction cannot, so it is asked for rather than assumed. It needs the structured engine — --no-docling turns it off and the run says so — and a model that is fetched once with docling-tools models download-hf-repo docling-project/CodeFormulaV2; where the model is missing the run refuses at the start rather than failing on the first page that holds a formula.

It is not free. Measured on one 20-page paper on this machine: 17.7 s without and 348.7 s with, for 56 transcribed formulas where the text layer yielded none. Prose gains nothing from it.

The transcription is appended under its own heading rather than replacing the text. On that paper the structured engine transcribed the formulas and lost 11% of the words, while native extraction covered the text and carried no formula — so choosing between them gives up one of the two. The record reports how many were transcribed as formulas.

Verification

Each conversion is compared against the text the original exposes. The index records the coverage achieved per file. Documents that expose no text — scanned drawings, for example — are marked unverifiable, since no text original exists to measure against.

The reference is read by a library independent of the layout and OCR engines, so for those the comparison is between two different readings of the document. For native extraction it is not: the reference and the extraction are the same read, and the comparison measures what the subsequent structuring and compaction kept. The reference is read once and used for both.

Packaging

mdcx pack --output ./Documents_md --target corpus.mdcx --key "passphrase"

--output accepts a folder of Markdown, a single file, or a .jsonl file with one record per line, which is how a catalogue of records is packed without first writing it to disk.

--fast selects a lower compression preset, and --preset names one directly. Sealing is a fixed price paid on every write: a package meant to be distributed is compressed once and downloaded many times, while a corpus rebuilt whenever it grows pays the clock and not the bytes. The level travels inside the compressed stream, so a package written at any level is opened by any reader, and nothing else about the package changes.

--compression chooses what compresses the body. The header records the name and a package is read by what its header names, so existing packages are unaffected and either kind is opened by a reader that has the module for it. LZMA is the default and needs nothing; zstd needs mdcx[zstd]. Which is smaller depends on the material — measured on one corpus zstd was smaller, on another larger — while zstd is consistently faster to read back, which is what a server pays before it can answer anything.

pack returns seconds_index and seconds_seal, and seconds_index_by_phase divides the first among reading the folder, cutting passages, counting terms, building the shape index and encoding. With --multilingual the encoding dominates the rest by two orders of magnitude, and what it costs is set by how long the passages are rather than how many: the model is charged per token.

The figure to plan a large corpus with is therefore a rate and the passage length it was measured at. On one 6 GB card, encoding 600 passages of each length:

A rate quoted without its passage length is not usable, and measuring one against text generated for the purpose gives the machine rather than the work: a consumer comparing 440 passages/s from invented 130-character texts against 25 from their own 942-character passages read the difference as a defect in this library. The phase figures are what settle it for a given corpus.

Querying

mdcx search corpus.mdcx "the question" --key "passphrase"
mdcx info corpus.mdcx
mdcx export corpus.mdcx --target ./restored --key "passphrase"

info reads the package header and requires no key: format, issuer, creation date, document and passage counts, integrity, signature, language, dates and calibration.

--prefer recent enters as a third ranking fused by rank rather than as a decay applied to a score. It orders rather than filters: an older document that answers better is still returned. Where both engines agree on the best passage the date does not move it; where they disagree, it decides. When the preference cannot be applied — no meaning index, or no dated passage in the answer — the reply states so rather than returning silently unchanged.

Retrieval model

Word matching and dense retrieval answer different questions and are kept side by side. The lexical index knows that a document contains a term; the dense index knows that a document means something similar. A query in one language cannot reach a document in another by word matching, because the words are not shared; representing meaning is what crosses that boundary.

The two are merged by reciprocal rank rather than by score. A BM25 score is unbounded and depends on the corpus it was measured in, and a cosine runs from zero to one and does not; the two cannot be added, and normalising them introduces a weighting nothing justifies. Rank is what both engines agree on.

Documents are ranked as documents rather than as isolated passages, so a document that answers in several places is not outranked by one that mentions the terms once.

Calibration

A reply states when nothing in the corpus bears on the question, rather than presenting its nearest passage as an answer. The threshold for that judgement is a property of the corpus and is measured when the package is built.

An absolute threshold cannot serve: how near a corpus comes to a question it answers depends on what the corpus contains, so a value set on one collection marks nothing on another. Each package therefore records its own reach, and the judgement is made against that.

pack --focus "<question>" takes the reach from the questions a package exists to answer rather than estimating it from the passages. It is the better source where passages do not resemble questions — a catalogue of abstracts, for example, where every record shares a rhetorical shape. Repeat the option to give several. info reports which of the two was used.

mdcx calibrate corpus.mdcx --key "passphrase" --question "..." --question "..."

calibrate measures a package that already exists, without its source material: the measurement needs the vectors, which are in the package, and questions, which come from outside. The package is rewritten around the changed measurement — same documents, same passages, same vectors — and the provenance is recorded as focus-after rather than focus, since measuring while packing and measuring afterwards describe the corpus at different moments. A signed package requires its signing key.

A package built before this measurement existed carries none, and is judged by the previous thresholds unchanged.

Vocabulary

archive.vocabulary(connection)           # term frequencies, and the rule
archive.unknown_terms(connection, text)  # terms this corpus has never seen
archive.unfamiliar(connection, text)     # the same, with the share and a flag

The index records terms of three characters or more, and one character where the writing system does not separate words. Absence from the table therefore has two meanings — the corpus never saw the term, or the index would never have recorded it — and a caller weighting terms by rarity must distinguish them, since an absent term takes the maximum weight. unknown_terms applies the rule.

Function words are not removed on top of that. Which words carry no information depends on the question, and die in die casting or les in Les Misérables carry meaning in the language being searched. What mdcx can state is what it never recorded.

The keys are normalised — folded case, folded accents — so a caller tokenising with search.tokenize_text must normalise before looking a term up. unknown_terms does so.

The word index does not cross languages, while the meaning index does. Asked across languages, unknown_terms returns every term of the text, which describes the corpus's language rather than its contents. unfamiliar reports the share and a cross_language flag; prefer it where the language of the question is not known to match the corpus.

Two signals, and their disagreement

archive.assess(connection, text) returns both.

A cosine cannot distinguish the senses of a homonym, because a multilingual embedding places them together: a corpus of algebra can come close to a question about graph colouring and return passages about the plot of a function. The literal vocabulary does distinguish them, since the word the question turns on is absent from it.

assess reports both signals and overrules neither. Whether an unfamiliar word should refuse a query depends on what the query is for, and a word may be peripheral to it.

Transcription recovery

A word that optical character recognition transcribed wrongly is a word the index does not contain, and literal matching cannot return it. The passage remains in the corpus and the question cannot reach it; the reply is empty, with no error to indicate why.

Optical recognition confuses characters that are drawn alike. Grouping characters by shape therefore places a misread word and its original in the same bucket. The bucket is too coarse to answer with, so it is used as a filter: it proposes candidates, and edit distance decides among them.

mdcx pack --output docs --target corpus.mdcx --key "..." --shapes
mdcx shapes corpus.mdcx --key "..."      # for a package written without it

The index is optional because it enlarges the package and serves only corpora that passed through optical recognition.

Three limits apply. The filter addresses errors of transcription and not of typing — its advantage comes from the shape of the characters, and so does its boundary. A substantial share of what it proposes is not a transcription error, since for Latin script the shape classes are coarse. And it operates only where no term of the query matched, since one term the corpus does contain is sufficient to return passages.

Where a passage was located under a different spelling the reply records it, in read_as. Presenting such a passage as a literal match would attribute to the document a word it does not contain.

The technique is that of character shape codes, described by Spitz at Xerox in the 1990s and applied there over document images to avoid a full recognition pass. Here it is applied over text already converted, for the errors the conversion left behind.

Attachments

A document with indexed: false in its front matter is kept in the package — signed, encrypted, in the same file — and excluded from the index, the vectors and the passage count. It is intended for material that accompanies a corpus without being text to search: a certificate, a table of coordinates.

export restores it verbatim, and the MCP document tool returns it by name, which is the only way to reach it since it cannot be found by searching.

pack reports which document contributed the most passages, and says so when one contributes a third or more.

Dates

A package records when each work is from, together with how that was determined.

mtime is never used unless --date-from-mtime requests it, since it is available for every file and describes the file rather than the work. Where nothing reliable is found the date is absent.

The copyright year printed in a document is deliberately not used: a reprint carries the year of the printing rather than of the edition.

mdcx pack --output docs --target corpus.mdcx --key "..." --dates dates.csv

Each line of the sidecar is path,date[,provenance]. info reports how many documents carry a date and the span they cover, and every passage in a reply carries dated and dated_from.

Correspondence

Where a collection is correspondence, pack records which side each document came from, and search --only received|sent restricts the query to one of them. The classification is taken from the directory structure of the input.

Incremental packaging

mdcx pack --output ./docs --target corpus-2.mdcx --key "..." \
    --multilingual --reuse corpus-1.mdcx

A passage whose text has not changed has the same vector. --reuse reads the vectors of an existing package and encodes only what is new. The vectors are read from a package already encrypted with the same key; no intermediate store is created, since a vector permits the text it represents to be approximated.

Reuse also carries the calibration forward. A package given its questions with --focus passes them to the package built from it, so a corpus that grows retains the threshold it was calibrated with. Passing --focus again overrides what was inherited, and the summary records when a calibration was inherited.

Compression and encryption are properties of the whole file, so they cost the same whether one document was added or the corpus rebuilt. pack reports seconds_seal so a caller writing frequently can decide how often to write.

The database is built in the file it will be written to, and the documents are read into it one at a time; it is then compressed, encrypted and hashed in one pass, in blocks. The peak memory of writing a package is therefore set by the block rather than by the size of the corpus — measured over a corpus and one four times larger, the peak rose by two per cent.

Resident conversion

The conversion engines load once per process. A caller running one process per document — the usual way to distribute work and isolate failures — pays that load for every document.

from mdcx.convert import resident

with resident.warm(report=print) as convert:
    for path in queue:
        record = convert(path, output_root)

The files written are those mdcx-convert writes for that document, chapters included: a long PDF becomes one Markdown per chapter in a folder of the document's name, alongside an index Markdown linking them. split=False asks for the document as a single file instead.

record is the record mdcx-convert writes. Where the document was split it is the index record, and the chapter records are under record["chapter_records"]. resident.convert_documents(paths, output_root) is the same as a generator, yielding one record per document, and resident.cost_of_starting() returns the seconds spent, so a log can report them. Several such processes each amortise their own startup.

Handing a whole folder to mdcx-convert also amortises the load, at the cost of the ordering, per-document handling and failure isolation a queue provides.

Expressions

A lexical index is built out of words, and the rule that decides what a word is discards the symbols. For prose that is correct. For a corpus interrogated by statement it removes the content: pq | b(b+p+q) and pq | b(b-p-q) differ by one sign and reduce to the same terms, and asked on their own they reduce to no terms at all and retrieve nothing.

--expressions keeps them beside the words, as their own index. A token that mixes symbols with alphanumerics is an expression; prose yields none. A question carrying an expression the corpus states is then answered by it, and one carrying an expression the corpus does not state is answered with nothing rather than with the passages that merely discuss the subject. Lookup is exact, since anything looser returns the confusion the index exists to remove.

An expression containing a space is not recovered: deciding where a formula ends inside a sentence is a different problem, and guessing would fill the index with fragments of prose.

Quotations across a cut

A quoted phrase often starts in one passage and ends in the next. --quotes document, the default, keeps the normalised text of every document and looks in it: there is no limit to how long a quotation may be, and it is a second copy of the corpus — measured on one package, a third of the file.

--quotes boundary indexes the join between consecutive passages instead, in a table that keeps the index and not the text. A quotation spanning one cut is found by phrase match; one longer than the join is not, which the whole copy would still find. The header records which shape a package has.

MCP server

MDCX_FILE=/path/to/corpus.mdcx
MDCX_KEY=passphrase

python -m mdcx.mcp_server

The key is supplied through the environment rather than on the command line, where it would be visible in the process table.

MDCX_FILE and MDCX_KEY accept several packages, separated by the platform's path separator or by commas, queried as one corpus. One key serves all of them, or one key per package in the same order.

Three tools are exposed:

A search reply carries the passages with their source and rank, and reports what it could not do: warning when nothing in the corpus bears on the question, read_as when a passage was located under a different spelling, prefer_applied when a requested preference could not be honoured, and unknown_terms naming, per package, words of the question that package has never seen.

Where several packages are served, similarity is computed over all of them together while answerable_at is the lowest of their calibrated reaches, so the threshold in force comes from the narrowest package rather than from the one a passage came from. answerable_at_by_package reports each.

The server ends its own process when its client disconnects, and releases the embedding model after MDCX_IDLE_UNLOAD_MINUTES of inactivity — 30 by default, 0 to disable. The model reloads on the next question.

Sources

mdcx converts, packages and answers. It does not fetch, and depends on no network of its own. A catalogue is a plugin, declared through the mdcx.sources entry point group and meeting the contract in mdcx.sources:

[project.entry-points."mdcx.sources"]
oapen = "my_package.oapen:Catalogue"

Keeping adapters out is deliberate. What appears to be a simple HTTP client is not: catalogues differ in how a download is reached, how paging is expressed, and how identifiers are formed, and that knowledge belongs with whoever holds it. Answering questions over an existing package requires none of this.

What does not depend on any particular catalogue is provided:

python -m mdcx.sources --check <name>

sources.looks_like(data, "pdf")   # by signature, not by declared type
sources.identify(data)            # what the bytes are instead
sources.patiently(call)           # retries RateLimited, honours Retry-After
sources.conforms(source)          # the conformance check, as a function

Candidate.download records which server a file would be fetched from, where the catalogue knows it without a further request, and Candidate.host reads it back. The conformance check reports when one host accounts for more than half of the candidates.

tests/test_sources_kit.py contains a reference source of about twenty lines, implemented against no network.

Portable paths

Every document carries a portable path of the form @/folder/document.md, relative to the root of the collection. Absolute paths are not stored, so a package is readable on a machine whose directory layout differs. export reconstructs the directory structure from them.

Signing

mdcx keygen
mdcx pack --output ./docs --target corpus.mdcx --key "..." \
    --issuer "Organisation" --signing-key <hex private key>
mdcx verify corpus.mdcx --public-key <hex public key>

Signing is Ed25519 over the digest of the encrypted body together with the header, so altering either invalidates the signature. Verification requires only the public key, not the passphrase.

Encryption

The package body is encrypted with AES-256-GCM, with the key derived by scrypt from the passphrase and a per-package salt. The header — format, issuer, counts, integrity digest, signature, language, dates and calibration — is outside the encrypted body and readable without the key, so a recipient can decide whether a package is worth opening.

An empty passphrase is refused at packing time. Supply the passphrase through MDCX_KEY, --key-file, or --key - to read it from standard input; --key places it in the process table for the duration of the command.

Package format

A .mdcx file is a magic number, a JSON header, and an encrypted body. The body is a compressed SQLite database holding:

Embeddings are stored in half precision and renormalised when read. A package written by an earlier version omits the tables added since, and every entry point that reads one checks first, so an older package continues to answer.

Limitations

• Conversion fidelity is measured against the text a document exposes. A document that exposes none cannot be verified, and is marked as such.
• Retrieval across languages requires the multilingual extra and its model.
• A package written before 1.27.0 carries the tokenizer it was built with, which cut scripts that write their vowels as combining marks at every mark. Such a package opens and answers unchanged; rebuilding it indexes those words whole.
• The transcription-recovery filter addresses errors of transcription, not of typing, and a substantial share of what it proposes is not an error.
• A package is decrypted into memory in full. A corpus larger than available memory is held as several packages queried as one, and the server opens each only when a query reaches it; info reports which are open and what they hold. Writing a package does not have this bound: it is built in the target file and sealed in blocks.
• Serving several packages, the lexical scores are computed against the packages taken together, so they can be ordered against one another. A caller using archive.query directly on several packages should pass the same corpus to all of them — corpus_statistics_over gathers it — or the scores are on as many scales as there are packages.
• Calibration thresholds are measured per corpus. A package built before that measurement existed is judged by fixed thresholds.

Tests

pytest

Contributing

See CONTRIBUTING.md. Issue templates for defects and measurements are under .github/ISSUE_TEMPLATE.

Security

See SECURITY.md for supported versions and how to report a vulnerability.

Releases

Released to PyPI, to the MCP registry as io.github.jorgell23-sys/markdown-document-search, and archived on Zenodo.

CHANGELOG.md records what changed in each release.

Authorship

Jorge Ellena G. See NOTICE for third-party components and their licences.

Citation

To cite mdcx in academic work, use CITATION.cff or the DOI:

https://doi.org/10.5281/zenodo.22015991

Licence

Apache License 2.0. See LICENSE.