For AI agents: the complete documentation index is at https://docs.unstract.com/unstract/llms.txt. Every docs page is also available as Markdown: drop the trailing / from its URL and add .md (the home page is https://docs.unstract.com/unstract/index.md).
Use a capable VLM as the advanced LLM (GPT-4o, Claude Sonnet, Gemini Pro). Table extraction accuracy depends heavily on the model's vision and reasoning capabilities.
Use a fast, inexpensive model as the lite LLM (GPT-4o-mini, Claude Haiku, Gemini Flash). Agent 1 makes one call per page, so cost adds up on large documents.
Ensure your X2Text adapter produces clean text. LLMWhisperer is recommended for PDFs with complex layouts.
Focus the target table description on what makes the table unique — its title, column headers, location in the document, or the type of data it contains.
Test with a few representative documents before processing in bulk.
When your document has multiple tables, clearly distinguish the target table from the others.
Keep the JSON structure simple and flat when possible. Deeply nested schemas are harder for the code generator to map correctly.
Use clear, descriptive field names that relate naturally to the table columns.
For grouped output, define a nested schema where the group-level fields sit alongside an array of child items. The extractor infers grouping from the schema shape — no separate setting is needed.
When using summary totals, place the summary object at the same level as the child items array so the extractor can compute aggregate values per group.
Test the JSON structure on a few documents before using it in production pipelines.
The extractor automatically merges rows that span page breaks. If you notice split rows, verify that the row identifiers (first column values) are consistent across pages.
For very long tables (100+ rows), consider whether all rows are needed or if the page range can be narrowed.
VLM-dependent accuracy. Extraction quality depends on the advanced LLM's vision capabilities. Some models handle complex table layouts better than others.
Complex merged cells. Tables with extensively merged cells (spanning multiple rows and columns) can reduce extraction accuracy.
Handwritten tables. Handwritten or heavily stylized tables are not well-supported. The pipeline works best with printed, digital-native tables.
Sandbox restrictions. The code executor (Agent 5) has limited module access. Transformations requiring specialized libraries (e.g., pandas, numpy) are not supported.
60-second timeout. Generated transformation code must complete within 60 seconds. Very large tables with complex transformations may exceed this limit.
Cost. The multi-agent pipeline makes multiple LLM calls per document. For simple tables, standard table extraction (enforce_type: table) is more cost-effective.