Skip to main content
Compares two or more already-extracted sides and returns a structured match matrix. Display name “Compare”. On by default for new agents. This is the default for PO matching, invoice-vs-PO, 3-way match, or any “do these documents agree” job. Extract each side first (structured_data_extraction, or an ERP read). Then call compare. Do not use ai_call or python_code_tool just to diff extracted fields. decide_next_action writes prompt every run. Two valid patterns:
  1. Compose the prompt from prior step results already in history — both sides, tolerances, match rules — in citation form.
  2. Reference extraction variables in the prompt (Invoice:\n${VAR_2}\nPO:\n${VAR_3}). The platform substitutes those tokens before the compare LLM runs.
The compare LLM does not see earlier chat. Put all required data in prompt. Convert extracted fields to citation form:
  • {"quantity": {"value": 2, "word_id_groups": [7]}} → Quantity: 2[cite:7]
  • Never write word_id_groups=[7] in the prompt.

​
Authentication and enablement

No integration. Uses the agent’s LLM. Default-on. Applied learning rules for this comparison are injected by the orchestrator.

​
Inputs

  • prompt (required): the comparison, including both sides, match rules, and citation markers. May embed ${VAR_N} references to prior extracted_data variables.

​
Output

Primary output is structuredContent — a JSON comparison matrix, not a prose verdict.
  • Fields — header rows (PO number, vendor, totals) with per-document values and _row_meta.compare (status + reason).
  • LineItem1 … LineItemN — one array per logical line. Each array is row-per-document with a Source field (Invoice, PO, GRN). Do not treat this as column-per-document.
  • _group_meta — per-line status, mismatch_columns, and per-field verdicts under fields.
  • _doc_meta — title, documents[] (name, icon, priority), and comparison_summary.
Status values: match, mismatch, missing. Every verdict includes a short reason. When a learned rule drove a verdict, the object also has rule_id (rule-N). Step metadata holds provider, model, token counts, and applied_memories. The task feed renders this matrix (CompareFeedItem) with match/mismatch badges and citation highlights when word_id_groups are present.

​
PO / 3-way matching

  1. Extract the invoice with structured_data_extraction.
  2. Extract the PO file, or read it from the ERP (read_sap_purchase_order, …). For 3-way, also extract/read the GRN.
  3. Call compare with a prompt that includes every side in citation form (composed by decide_next_action, or referenced via ${VAR_N}).
  4. Read the structured matrix for later review or writeback. Needing line-level JSON is not a reason to replace this tool with ai_call.

​
Not a substitute for

  • structured_data_extraction — extract fields first; compare does not read raw files.
  • llm_match — fuzzy match one value against a synced catalog (vendor/item), not document-vs-document.
  • ai_call — generic enrichment (coding, mapping, summaries). Not invoice-vs-PO.
Compare has no configured-tool bindings. An optional display-name alias still executes this native tool; call the base name compare if learned-rule forwarding should apply.

​
Limits and side effects

  • One LLM call per invocation. No external writes.
  • Large prompts are accepted (MaxTokens is high); billed per call plus tokens.

​
Expected errors

  • Empty prompt.
  • Provider failure.