Unstructured Transform MCP prompting strategies

After you connect your AI tool to the Transform MCP server, crafting effective prompts is key to getting the best results from Transform. Use the prompt strategies outlined below to improve the results you get from Transform. Use them as written, or adapt them to your own files and use cases. For the full set of options you can combine into your own prompts, see File parsing options.

Get high-fidelity result from one-shot prompts

If you want the highest-fidelity output from a single request, ask Transform to try multiple partitioning strategies and synthesize the best result. This works best when your AI tool or agent is backed by a strong vision-language model (VLM) model:

Transform this document into as high fidelity of an output as possible by using 
the top contending parsing strategies and then analyze and synthesize the best 
results into a single output.

This prompt can produce worse results if your AI tool or agent is backed by a weak vision-language model (VLM). A weak VLM may use its own flawed interpretation of the document, rather than favoring the stronger partitioning strategy outputs.

Chain prompts for incremental experimentation

Instead of parsing, enriching, chunking, and embedding your files in a single request, you can chain requests together. The output from one request becomes the input for the next. Chaining is useful when you want to experiment with different options at each stage without repeating earlier stages. The following three example prompts chain together to take a document from initial parsing, through enrichment, to a retrieval-augmented generation (RAG)-ready output.

Step 1: Perform the initial file parsing

Start with parsing the file into JSON using the High Res partitioning strategy. Ask Transform what other options are available so you can experiment further:

Transform this document into JSON using the hi-res parsing strategy. 
For future experimentation, let me know what options are available for 
parsing strategies and their settings.

Step 2: Add enrichments to improve results

Using the JSON output from step 1 as input, add a few typical enrichments. Ask Transform what other enrichment options are available:

Enrichments are only available if you used the hi-res partitioning strategy in the previous step. The VLM partitioning strategy already includes many of the same enrichment capabilities built in. The fast partitioning strategy does not support enrichments. For more information, see Partitioning.

For more information, see Enrichments.

Using the previous JSON output as input, add the image enrichment, 
table to html enrichment, and generative OCR enrichment. For future 
experimentation, let me know what options are available for enrichments 
and their settings.

Step 3: Prepare files for RAG

Using the enriched JSON output from step 2 as input, prepare the document for retrieval-augmented generation (RAG) by chunking it and adding embeddings. Ask Transform what other chunking and embedding options are available:

Chunking divides the text-based content in documents into manageable “chunks” of roughly the same size, in chronological order, to stay within the limits of an embedding model and to improve retrieval precision.
Embedding generates vectors that represent the text extracted by Unstructured, and stores them next to the text itself.

For more information, see Chunking and Embedding.

Using the previous JSON output as input, get the document ready for 
RAG, by chunking the document and adding embeddings using the default 
settings. For future experimentation, let me know what options are 
available for both chunking algorithms and settings as well as 
embedding models.

Next steps

Control Transform file parsing output: Control how the Unstructured Transform MCP server instructs Transform to partition, enrich, chunk, and embed the data based on 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.

Questions? Need help?

For technical support, request support.

​Get high-fidelity result from one-shot prompts

​Chain prompts for incremental experimentation

​Step 1: Perform the initial file parsing

​Step 2: Add enrichments to improve results

​Step 3: Prepare files for RAG

​Next steps

​Questions? Need help?

Get high-fidelity result from one-shot prompts

Chain prompts for incremental experimentation

Step 1: Perform the initial file parsing

Step 2: Add enrichments to improve results

Step 3: Prepare files for RAG

Next steps

Questions? Need help?