mdcx

Convert a document collection to verified Markdown, package it into a single encrypted file, and make it queryable by agents through the Model Context Protocol.

The problem

An agent answering questions about a document collection has two options. It can receive the documents in its context window, which is expensive and bounded by the window size. Or it can query a component that already knows where each item is.

Measuring one specific query — where the minimum pipe diameter to be modelled in 3D is stated — over a real collection of 99 documents and 180 MB, using the cl100k_base tokenizer:

The 435 comprise 20 for the question, 274 for the retrieved passage and 141 for the answer.

The first row costs the entire collection for a concrete reason: a PDF cannot be searched, it is a binary, and without prior conversion there is no way to know which of the 99 documents holds the answer. They all have to be extracted and read.

This is one measurement, not an average: the saving depends on how much text an answer requires. What does not vary is the shape of the change. The work does not disappear, it moves from the context window — which is billed and finite — to the CPU, which is not. That is why the local column rises rather than falls.

The three stages

Conversion. Each document is converted to Markdown and checked against the text the original actually exposes, read with a library independent from the engine that performed the conversion. Content the structured engine omits is appended verbatim rather than reported as lost.

Over the collection used during development — 99 documents, 1,144,553 reference words — 594 words were not recovered, a global coverage of 99.948%. Of the 184 documents exposing text, 116 came out at exactly 100% and none below 99.5%. The remaining four are scanned drawings containing no text at all in the file: they were read by optical character recognition and are marked as unverifiable, because no text original exists to measure them against.

Packaging. The corpus, its search index and the provenance of every passage fit into a single .mdcx file, encrypted with AES-256-GCM, whose header can be read without the key. From 8.8 MB of Markdown to 3.9 MB in one file.

Retrieval. A query returns the passages that answer it with their exact source. Over the 20 real queries used for tuning, the correct document appears within the top five results in 19 cases and within the top ten in all 20.

Installation

The package separates querying from conversion, because they have very different requirements.

Conversion is what pulls in the heavy dependencies. Someone who receives an .mdcx file and only needs to query it installs neither Docling nor PyTorch.

The multilingual extra is what makes a query in one language reach documents written in another, and most of its size is the model, downloaded once on first use. A corpus written in a single language does not need it.

Converting a collection

pip install "mdcx[convert]"
mdcx-convert --input ./Documents --output ./Documents_md

The output mirrors the input directory structure, adds a global index, and records for each file the coverage achieved against its original.

Packaging and querying

mdcx pack --output ./Documents_md --target corpus.mdcx --key "..."
mdcx info corpus.mdcx
mdcx search corpus.mdcx "where is the minimum diameter stated" --key "..."
mdcx export corpus.mdcx --target ./restored --key "..."

info reads the header without the key, so the issuer and the integrity of a file can be checked before opening it. export rebuilds the original folder: a format that cannot be left is a trap, however well intended.

Using it as an MCP server

The server requires Python and this package. It does not require the conversion stack, so the footprint is about 50 MB.

{
  "mcpServers": {
    "mdcx": {
      "command": "python",
      "args": ["-m", "mdcx.mcp_server"],
      "env": {
        "MDCX_FILE": "/path/to/corpus.mdcx",
        "MDCX_KEY": "package-key"
      }
    }
  }
}

Alternatively, with uv the server runs without a prior installation, which is the usual arrangement for Python MCP servers:

{
  "mcpServers": {
    "mdcx": {
      "command": "uvx",
      "args": ["--from", "mdcx[mcp]", "python", "-m", "mdcx.mcp_server"],
      "env": {
        "MDCX_FILE": "/path/to/corpus.mdcx",
        "MDCX_KEY": "package-key"
      }
    }
  }
}

Three tools are exposed. search returns the passages answering a question, each with its source document and portable path. info describes the corpus and the fidelity of its conversion. document returns a full document when passages are not enough.

The server verifies the package before it starts listening, so a wrong path or key is reported immediately rather than on the first query.

Tests

pip install pytest
python -m pytest tests/ -v

Eighty tests run in three groups.

test_stress.py covers hostile inputs: empty and corrupted files, names in other alphabets, malformed queries including SQL injection attempts, truncated and tampered packages, concurrent access, and compaction against content loss.

test_languages.py measures retrieval itself. It builds a corpus of 136 documents written in 34 languages across 11 writing systems, and checks two properties: that a query written in a language retrieves the documents written in that language, and that a term shared by several languages returns the documents of all of them. The second is what keeps a search from narrowing to one language, whether the language of the query or that of the corpus.

test_multilingual.py measures retrieval across languages: that a query written in one language returns documents written in others, that every writing system is reachable, and that merging the two engines does not cost the precision the lexical engine has on its own. These are skipped when the multilingual extra is absent, which is a supported configuration rather than a failure.

Languages

Retrieval is lexical and script-aware. A query matches the words that appear in the documents, in whatever writing system they were written, and the results of one search may come from documents in several different languages at once. The language of a corpus is recorded and reported by info; it describes the corpus and never restricts what a query returns.

The following are verified by tests/test_languages.py, which builds one corpus holding the same four subjects — algebra, botany, printing and baking — written in every language listed, and then issues a query in each language. Every query competes against the three other documents in its own language and against the whole of the rest of the corpus. All 136 queries return the expected document in first place.

Support is a property of the writing system rather than of the language, so a language written in one of these scripts is covered whether or not it appears in the table. Three properties make that hold:

Words are runs of letters, digits and the combining marks that belong to them. The marks are read from the Unicode database rather than listed, which is what keeps the vowels of Devanagari, Bengali, Tamil and Thai attached to the letters they modify instead of splitting each word into fragments.

Accent folding is limited to the combining marks that represent an accent placed on a letter, so that café matches cafe. The vowel signs of Indic scripts and the points of Hebrew and Arabic are left in place, because there they carry the sound of the syllable rather than decorate it.

Writing systems that do not separate words with spaces — Chinese, Japanese, Korean, Thai, Lao, Khmer, Burmese, Tibetan and Javanese — are indexed by character, at query time and at index time alike. This is what a lexical index can match without a segmenter trained on one particular language, which would serve that language and leave every other one where it started.

Everything above concerns matching by word, which is matching within a language: the words of a query have to be present in the document. A query written in one language reaches a document written in another only where the two share a term, as proper names and loanwords often do. When a query returns nothing and none of its terms appear in the index, the result says so and names the language of the corpus, so an empty answer can be told apart from material the corpus does not hold.

Crossing between languages is a separate capability, described in the next section. It is optional because it needs a model, and it is merged with word matching rather than replacing it.

Searching across languages

Word matching is matching within a language. A Spanish query and a German document about the same subject share no word, so there is nothing for an index of words to find, however well the words are tokenised. Measured on a corpus written in thirty-four languages, a query in one language retrieves 4.2 per cent of the documents on its subject: essentially only the ones written in the language of the query.

Reaching the rest requires representing meaning rather than spelling. A multilingual embedding model places a sentence and its translation near each other, so a document can be found through what it says instead of the words it happens to use. Built with --multilingual, a package stores a vector for each passage next to the passage itself, under the same encryption, and the same query then retrieves 96.9 per cent of them.

pip install "mdcx[multilingual]"
python -m mdcx.archive pack --output ./corpus_md --target corpus.mdcx \
    --key "..." --multilingual

The corpus is encoded once, when the package is built. Whoever receives it encodes only their own queries.

Both engines, merged

The two engines are kept side by side because they fail in opposite directions. On the same corpus:

Meaning alone reaches almost everything and loses precision on the language of the query, where an exact word is exactly what should decide. Words alone are precise and cannot leave that language. Merged by reciprocal rank, each covers what the other cannot, at the cost of one document out of a hundred and thirty six against the lexical engine on its own ground.

Rank is merged rather than score because the two scales have no common meaning: a BM25 score of 8 and a cosine similarity of 0.8 cannot be added, and normalising them introduces a weighting that nothing justifies.

--mode lexical and --mode semantic select a single engine when the comparison matters.

Choosing the model

The default is BAAI/bge-m3, chosen by measurement on FLORES-200, a corpus of sentences translated by professionals into two hundred languages. The task is to find a sentence given its translation, among candidates from the same corpus, in both directions of every language pair.

What decides is not the average but the pair of languages the model handles worst. A model that averages well while collapsing on one language does not serve whoever reads in that language.

Measured over 132 language directions covering ten writing systems, with 50 candidates per query. On a larger run of 1 122 directions across all thirty-four languages with 100 candidates, LaBSE scored 99.6% mean and 93.5% on its worst language, with no pair below 90%.

Another model can be used by name:

MDCX_MODEL=sentence-transformers/LaBSE python -m mdcx.archive pack ...

A package records the model that encoded it. A query encoded by a different model lands elsewhere in the vector space, so the mismatch disables meaning rather than returning results that look ranked and are not.

What this does not do

Retrieval finds documents that say something close to the query. It does not translate them: passages are returned in the language they were written in. Nor does it make an unrelated document relevant because the model recognised the subject; the ranking still has to place it, and the measurements above are what it does place.

Paths

No output contains absolute paths. Every document is identified by a pseudopath beginning with @/, resolved against the folder or package containing it, so a corpus remains valid wherever it is stored: local disk, network share or cloud.

Signing

A package can be signed so that its issuer can be proven rather than merely declared. The signature covers the digest of the encrypted body, so it attests both origin and content, and is verified without the encryption key.

mdcx keygen
mdcx pack --output ./Documents_md --target corpus.mdcx --key "..." \
          --issuer "Acme Ltd" --signing-key <private-key>
mdcx verify corpus.mdcx --public-key <public-key>

Verification also requires the body to be intact: a signature covering only the recorded digest would otherwise accept a package whose contents had been replaced while its header was left untouched.

The issuer field alone is free text and proves nothing. Only a signature does.

Encryption

The package encrypts at rest and decrypts in memory when opened; nothing is written to disk in clear. This protects a file in transit. It is not the same as searching over encrypted data without ever decrypting it, which is a separate field with documented leakage attacks and per-query costs measured in seconds.

The key is derived with scrypt, which makes guessing slow: about 8 attempts per second, each requiring 32 MB of memory, which prevents parallelisation on a GPU. Even so, the real strength is the passphrase: a dictionary password falls in a day.

Authorship

Conceived and directed by Jorge Ellena G., programmed with the assistance of Claude (Anthropic).

Every decision in this package was made against measurements rather than convention: which conversion engine to use, which licence permits which, how to rank a search, which optimisations to accept and which to discard. Several were discarded precisely because they were measured — reducing the search candidate pool appeared to be ten times faster and in fact lowered accuracy from 19 to 17 out of 20 — and those measurements are recorded alongside the decisions they justify.

Citation

Archived on Zenodo with a permanent identifier. The concept DOI always resolves to the latest version:

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

Licence

Apache 2.0. The software may be used, modified and sold, provided the copyright notice is retained.

PyMuPDF was deliberately avoided: its AGPL licence would require anyone using this software to publish their own under AGPL, including those offering it only as a network service.