Unstructured Transform MCP file parsing options

Control how the Unstructured Transform MCP server instructs Transform to partition, enrich, chunk, and embed the data based on your files.

After you connect your AI tool to the Transform MCP server, to send your files to Transform for processing, you can give your AI tool or agent a simple prompt such as the following, depending on where your files are stored:

Somewhere on your local filesystem: Replace <local-paths-to-your-files> in the following prompt with the paths to your target files:

Use the Unstructured Transform MCP server to parse the files at <local-paths-to-your-files>

Within a project: For project-based AI sessions, you can use this simple prompt instead:

Use the Unstructured Transform MCP server to parse the attached files in the associated project.

Attached to a chat: If you have added your target files directly to an AI chat session, you can use this simple prompt instead:

Use the Unstructured Transform MCP server to parse the attached files.

Transform will parse your files by using its default file parsing options, as described in the next section. You can change any or all of these options.

Default file parsing options

By default, when the Transform MCP server parses your files:

• The Auto partitioning strategy is used. Learn about partitioning.
• No language bias for OCR is applied.
• All pages are parsed.
• No enrichments are generated. Learn about enrichments.
• No chunks are generated. Learn about chunking.
• No text embeddings are generated. Learn about embedding.

You can change any or all of these options, as described in the next section.

Specify file parsing options

To learn how to specify file parsing options, after you connect your AI tool to the Transform MCP server, ask your AI tool the following:

What are all of the available options that I can specify when asking the 
Unstructured Transform MCP server to parse my files? Give me an example 
prompt for each of these options.

The following core options are available:

Option	Values / format	Default	Example prompt
Partitioning strategy	auto, fast, hi_res, or vlm	auto	Partition using the hi_res strategy. Learn about partitioning.
Languages	ISO-639 language codes	No language bias	Bias for OCR for en, es, and fr.
Page ranges	One or more page ranges	All pages	Parse only pages 1 through 5 and page 10.
Enrichments (must specify hi_res partitioning first)	image_description, table_description, table_to_html, ner, and generative_ocr	No enrichments	Add the table_to_html and image_description enrichments. Learn about enrichments.
Chunking strategy	chunk_by_character, chunk_by_title, chunk_by_page, or chunk_by_similarity, with chunking options specified (see the next section)	No chunking	Split into chunks based on character counts. Learn about chunking.
Embedding	An available provider and model (see later on this page)	openai and text-embedding-3-small	Learn about embedding.
Output format	md, json, html, txt	md	Return the results as md.

Only the json output format produces Unstructured document elements and metadata.

For chunking, the following options are available. Learn about chunking.

Option	Applies to	Example prompt
combine_text_under_n_chars	chunk_by_title only	Combine multiple elements into the current chunk if possible, with a maximum chunk length of 800 characters.
max_characters	All	No chunk may ever exceed 800 characters.
new_after_n_chars	All but chunk_by_similarity	Close the current element after reaching 800 characters.
overlap	All but chunk_by_similarity	Take the last 100 characters of each chunk and add it to the beginning of its immediate next chunk only.
similarity_threshold	chunk_by_similarity only	Use a similarity threshold of 0.75.

For available embedding provider and model options, enter the following prompt:

For the Unstructured Transform MCP server, for embeddings, which available embedding provider and model combinations can I specify?

For each available model, list the number of dimensions that it supports. Use the format ([number] dimensions).

List the default embedding provider and model combination, if I do not otherwise specify one. Give me an example prompt phrase so that I can specify one of these combinations in an ideal way.

Learn about embedding.

File parsing prompt examples

Additional partitioning strategy prompt examples:

• Partition using the default auto strategy.
• Use fast partitioning.
• Use hi_res parsing.
• Use vlm parsing.

Additional enrichment prompt examples:

• Add image descriptions for figures and diagrams.
• Add table descriptions for the tables.
• Convert tables to structured HTML.
• Extract named entities and their relationships.
• Use generative OCR because the document has handwriting.

Additional chunking strategy prompt examples:

• Chunk by title, combining any sections under 200 characters into the next one.
• Split this into chunks of 1000 characters each, with a 100-character overlap between chunks.
• Chunk this by similarity, starting a new chunk whenever the similarity threshold drops below 0.75.
• Chunk by character, starting a new chunk after 1500 characters, but cap each chunk at 2000 characters maximum.

Additional embedding prompt examples:

• Embed this using OpenAI's text-embedding-3-small model.
• Embed these contracts using Voyage AI's voyage-law-2 model.
• Chunk this by title, then embed it with IBM's all-minilm-l6-v2 model.

Additional output format prompt examples:

• Return the parsed results as Markdown.
• Return the parsed results as JSON.
• Return the parsed results as HTML.
• Return the parsed results as plain text.

Complete example prompts:

• Partition in hi-res mode, optimizing for French and English. Enrich it with table-to-HTML and generative OCR. Chunk it by title, combining sections under 500 characters. Embed it with Voyage AI's voyage-3-large model. Give me the final output as JSON.
• Use hi-res partitioning on this contract, pages 1 through 5 only. Run named entity recognition to tag parties and dates. Chunk it by character with a max of 1200 characters and a 150-character overlap. Embed using Voyage AI's voyage-law-2 model. Return the result as Markdown.
• Partition this technical manual with the VLM strategy. Enrich it with image descriptions and table descriptions. Chunk it by page. Embed with OpenAI's text-embedding-3-large model. Output it as HTML.
• Partition these plain text files using the fast strategy. Skip enrichments. Chunk it by similarity with a threshold of 0.75. Embed it using OpenAI's text-embedding-3-small model. Give me plain text output.
• Partition these filings in hi-res mode, pages 1 through 30 only. The filings are in German. Enrich with table-to-HTML and table descriptions. Chunk it by title, capping chunks at 1000 characters. Embed it with Voyage AI's voyage-finance-2 model. Return the output as JSON, keeping full Base64 for image elements.

Next steps

• Transform quickstart: Install the Unstructured Transform MCP server, drag and drop files, and have Transform start producing partitioned and enriched data based on your files in minutes.
• Write effective Transform prompts: Get the best results from the Unstructured Transform MCP server in as few requests as possible using these prompting strategies.
• Control Transform structured data extraction: Control how the Unstructured Transform MCP server extracts and formats structured data from your files.
• Control Transform generated sample code: Control how the Unstructured Transform MCP server generates sample curl or Python code that demonstrates how to use Transform to partition, enrich, chunk, and embed the data based on your files.