[BUG MODEL]: mistral-ocr-latest structured annotation (strict json_schema) loops on number fields, truncating the JSON output

[adamlarkem-wq]

Model

mistral-medium-latest

Request Payload

---

Request payload

POST /v1/ocr (via client.ocr.process), model mistral-ocr-latest (also reproduced on mistral-ocr-4-0), structured document annotation with strict: true and number fields:

{
  "model": "mistral-ocr-latest",
  "document": {
    "type": "document_url",
    "document_url": "data:application/pdf;base64,<PDF_BASE64>"
  },
  "include_image_base64": true,
  "include_blocks": true,
  "document_annotation_prompt": "Extract the invoice fields into the schema. Numbers as plain decimals.",
  "document_annotation_format": {
    "type": "json_schema",
    "json_schema": {
      "name": "invoice",
      "strict": true,
      "schema": {
        "type": "object",
        "additionalProperties": false,
        "required": ["company", "invoice_number", "items"],
        "$defs": {
          "Item": {
            "type": "object",
            "additionalProperties": false,
            "properties": {
              "product_code":     {"anyOf": [{"type": "string"}, {"type": "null"}]},
              "description":       {"anyOf": [{"type": "string"}, {"type": "null"}]},
              "quantity":          {"anyOf": [{"type": "number"}, {"type": "null"}]},
              "unit_price":        {"anyOf": [{"type": "number"}, {"type": "null"}]},
              "line_total_price":  {"anyOf": [{"type": "number"}, {"type": "null"}]}
            }
          }
        },
        "properties": {
          "company":         {"type": "string"},
          "invoice_number":  {"type": "string"},
          "items":           {"type": "array", "items": {"$ref": "#/$defs/Item"}},
          "subtotal_amount": {"anyOf": [{"type": "number"}, {"type": "null"}]},
          "tax_amount":      {"anyOf": [{"type": "number"}, {"type": "null"}]},
          "total_amount":    {"anyOf": [{"type": "number"}, {"type": "null"}]}
        }
      }
    }
  }
}

Input: a single-page PDF invoice with decimal line amounts (attached: sample.pdf).

---

Output

Output

The returned document_annotation is truncated, invalid JSON: a number field is emitted as the correct value, then expands into its full float64 decimal representation followed by an endless run of digits, never closing the value (company/invoice anonymized; the runaway line_total_price is verbatim):

{
  "model": "mistral-ocr-latest",
  "document_annotation": "{\"company\": \"ACME ELECTRICAL\", \"invoice_number\": \"PO-000000\", \"items\": [{\"product_code\": \"99999", \"description\": \"LIGHT 1200MM\", \"line_total_price\": 1487.500000000000255795384873613816452026367187500000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000001402960000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000",
  "usage_info": {
    "pages_processed": 1,
    "doc_size_bytes": 45532
  }
}

Intended value was 1487.5. Output = 1487.5 + its float64 binary→decimal tail (...0255795384873613816...) + an unbounded digit run → the value and the rest of the document are never closed.

---

Expected Behavior

Each number field is emitted as a finite decimal and the response is well-formed JSON conforming to the schema (e.g. "line_total_price": 1487.5).

---

Additional Context

Environment: mistralai Python SDK, endpoint ocr.process document annotations. Models mistral-ocr-latest (= mistral-ocr-2505) and mistral-ocr-4-0 both affected.

Reproduction conditions:

• Reproduces only with strict: true. Same schema / model / prompt / PDF does not degenerate when strict is off appears to be a constrained-decoding interaction with the unbounded JSON number grammar.
• Requires number (float) fields in the schema and a PDF with decimal values.
• Intermittent and document-dependent: ~70% of runs on one specific invoice, ~0% on other invoices, ~10% across a set of 20. Same document + same request, re-running eventually triggers it.
• include_blocks / include_image_base64 have no effect.
• The truncation always lands inside a number field.

How we found it: extracting invoices at scale, some documents intermittently returned unparseable JSON. Re-running the exact same request on the same PDF reproduced it ~70% of the time on the worst document.

Reproduction note: the attached sample.pdf triggers it ~70% of runs, please re-run a few times if the first is clean. Happy to share more sample documents privately.

---

Suggested Solutions

No response