Docparser.dev

Documentation search

Find an answer

Start typing to search the documentation.
    Open portal

    01 MEASUREMENT NOTES

    Start with the method, not the percentage.

    These are early measurements from one human-reviewed document. They are useful for understanding the current system, not a general accuracy promise.

    Docparser Labs single-template benchmark · measured 2026-07-24

    BENCHMARK METHOD

    1. 01Freeze the document and expected field values.
    2. 02Have a human review the golden record.
    3. 03Compare extracted values field by field.
    4. 04Keep scope, date and exceptions beside the result.

    02 METHOD

    How we measured this

    We compared each returned value with a human-reviewed golden record. A field counts as a match only when its extracted value exactly matches that record for this benchmark run.

    Sample: n = 77 fields on one document.

    Date: 2026-07-24.

    Scope: one template, one document and the current Gemini Flash extraction path.

    03 RESULTS

    Two counts, one transparent distinction.

    The broader score includes every field in the response under review. The extraction-only score removes two envelope keys the model was not asked to extract.

    96.2%

    All reviewed fields matching the golden record

    Docparser Labs · n = 77 fields on one document · 2026-07-24

    98.7%

    Extraction fields matching the golden record

    Docparser Labs · same sample · 2026-07-24

    Excludes two response-envelope keys that the model was not asked to extract.

    13.0 s

    Median end-to-end extraction

    Docparser Labs single-template benchmark, Gemini Flash · n = 1 document · 2026-07-24

    p95 was 14.1 seconds in the same early benchmark run.

    04 GOLDEN RECORDS

    The record can be corrected without rewriting history.

    A reviewer later corrected a due-date value while creating the benchmark record. The golden was wrong before that human correction; the delivered output was not silently rewritten to match it. The corrected due date became the benchmark's ground truth, while the published measurement stays attached to its own run.

    05 MODEL COST

    Token arithmetic, not a pricing promise.

    The numbers below show a measured model expense for this run. They are our model cost, not your price.

    MEASURED RUN / GEMINI FLASH

    1,938 input × $0.30 per million input tokens 3,032 output × $2.50 per million output tokens
    $0.008161 → ₹0.79 at ₹96.61/USD Input and output token use are shown separately; the displayed INR amount uses the stated conversion rate.

    06 NOT YET PUBLISHED

    What this page does not claim.

    • Accuracy across a broad supplier or document population.
    • Latency percentiles from a production-scale sample.
    • Customer pricing, an uptime target or certification status.