Objective Coding

Objective coding is the process of adding or correcting factual metadata on documents -- things like the document date, title, author, document type, and language. Unlike legal coding (which involves judgment calls about relevance or privilege), objective coding focuses on verifiable information that can be read directly from the document.

This matters because metadata drives everything downstream: search filters, sort orders, chronologies, and production sets all depend on accurate, consistent metadata. When that metadata is missing or wrong, your review suffers.

When to use objective coding

• Missing metadata. Documents were loaded without key fields like date or author.
• Incomplete imports. Metadata extraction during processing captured some fields but missed others.
• Large volumes needing standardization. Thousands of documents have metadata in inconsistent formats.
• Chronologies needed. You need reliable dates to build a timeline, but the existing date field is unreliable.

How it differs from legal coding

The workflow

Objective coding in Claira follows a seven-step process. Each step builds on the previous one.

Step 1: Open your Claira workspace

Navigate to your case in Claira and open the workspace where you will run the objective coding scan.

Step 2: Confirm text extraction

Before Claira can analyze a document, it needs extracted text to work with. Verify that your documents have been processed and that text extraction completed successfully.

Step 3: Create or connect destination fields

You need fields in Nuix Discover to store Claira's output. We recommend creating dedicated AI fields rather than writing directly to your primary metadata fields. This gives you a chance to QC the results before committing them.

Example fields:

Connect these fields in Claira's workspace settings so scan results are written to the right place.

Step 4: Build and test your prompt on a single document

Start with one document to make sure your prompt produces the output you expect. Open a document in the single review view, enter your prompt, and check the result.

This prompt is intentionally strict: it asks for a single date in a specific format, with a clear fallback value (--) when no date can be determined. That structure makes QC and downstream processing much easier.

Step 5: Run a bulk scan

Once you are satisfied with your prompt on individual documents, run it across your full document set (or a targeted subset) using Claira's bulk scan feature.

Monitor the scan progress in the workspace. Larger document sets will take longer, but you can continue working while the scan runs.

Step 6: Quality control

This is the most important step. Do not skip it.

• Filter for fallback values. Search your destination field for the fallback value (e.g., --) to find documents where Claira could not extract the metadata. Review these manually.
• Spot-check results. Open 20-30 documents across different document types and compare Claira's output to the actual document content.
• Validate edge cases. Pay special attention to documents with unusual formats, multiple dates, or ambiguous authorship.

Step 7: Update primary metadata fields

Once you have QC'd the results and are confident in their accuracy, copy the values from your AI fields (e.g., OC Date) into your primary metadata fields (e.g., Document Date). This can be done through Nuix Discover's bulk field update tools.

Multi-Code mode (single pass, multiple fields)

Use Multi-Code when one scan must populate multiple outputs (for example: date, title, author, and document type).

• Configure up to 8 fields in the dedicated Multi-Code section.
• Click Reset fields (top right of the Multi-Code card) to clear the shared instruction, field prompts, and destination field selections, and restore 2 empty rows.
• In Prompt Lab, when Claira detects a multi-output prompt, click Use Multi-Code to convert it automatically:
  • only text that applies to every field — your reviewer role, the matter background, global handling rules — moves into the Multi-Code shared instruction,
  • each output becomes its own field prompt (2-8 parts), carrying the criteria, allowed values and fallback that decide that answer. If your prompt ends with an output-format block such as RELEVANCE: [YESREL/NOREL], Claira treats those labels as the names of the outputs and pulls the matching definitions out of the body of your prompt into each field prompt, rather than leaving them in the shared instruction.
  • rows are rebuilt from the converted parts, and each row is named after its output. A destination field you had already mapped is kept when the name still matches, and an unmapped row is pre-filled when a case field has exactly that name — otherwise the row is left for you to set.
• If any converted field prompt still looks too thin to decide an answer on its own, Claira names those fields in a banner at the top of the Multi-Code card. Review them before scanning.
• If the conversion cannot be completed — Claira is busy, or the prompt does not split into separate fields — nothing is changed: your prompt stays exactly as you wrote it in Prompt Lab, and the message offers Try again. Multi-Code is never opened on your behalf with a half-finished configuration.
• Use the Insert into selector as the default focus for the shared instruction or a field prompt row (all rows are listed, including those still being configured).
• Open View History or Quickstart (templates or prompt generator): Claira asks where to put the content—the shared instruction, any existing field row, or New field (until you reach 8 rows). If that destination already has text, choose append or replace next.
• In bulk scans, token usage is charged per document using the same rules as other bulk scans, with Multi-Code costs depending on the Scan as dropdown selection:
  • Text + Multi-Code: 2 tokens per document (1 base + 1 for Multi-Code), no matter how many fields are configured (1 to 8 fields all cost the same).
  • Image + Multi-Code: 6 tokens per document (5 base + 1), regardless of how many fields are configured.
  • Audio + Multi-Code: 11 tokens per document (10 base + 1), regardless of how many fields are configured.
  • Video + Multi-Code: 21 tokens per document (20 base + 1), regardless of how many fields are configured.
  • Image, audio, and video modes are available on every plan. See Media scans for accepted file formats per mode.

Empty values (when a field has no answer)

Sometimes a document simply does not contain what a field asks for — an email has no contract date, a photo has no author. When that happens, Claira reports the field as Empty instead of inventing a value:

• In the single-document response card and the bulk-scan live feed, the field shows a muted Empty marker next to its name.
• Nothing is written to the destination field in Nuix Discover — any value already coded there is left unchanged. Claira never overwrites existing coding with a blank.
• A document where every field comes back empty still completes normally: it was scanned, and "nothing to extract" is a valid result. In bulk scans it appears in the live feed with its fields marked Empty and adds to the completed count, not the error count. It is billed like any other scanned document.
• Empty is not an error. Errors — an unreadable AI response, or a value that could not be written to Nuix Discover — are still reported as failed documents.
• If you prefer fields to stay untouched, rely on empty values rather than prompt-level fallback text: a fallback value like N/A is written into the field, while an empty answer leaves it unchanged.

Best practices

• Be specific in your prompts. "What is the date?" is too vague. "Identify the primary document date. Respond ONLY MM/DD/YYYY." is clear and testable.
• Group similar document types. If your collection includes contracts, emails, and memos, consider running separate prompts tuned to each type. Contracts have "Effective Date" while emails have "Sent Date" -- one prompt may not handle both well.
• Always QC before committing. The AI fields exist to give you a safety net. Use it.
• Use fallback values. A clear fallback like -- or N/A is far better than a blank field. It tells you Claira tried and could not find the answer, which is different from Claira not having processed the document at all.

Limitations

• Text mode reads extracted text only. In Text mode, Claira cannot read images, handwritten notes, audio, video, or content embedded in non-text formats unless they have been OCR'd or transcribed. For those documents, switch the Scan as dropdown to Image, Audio, or Video so the source file is sent directly to a multimodal model.
• OCR quality matters in Text mode. If the OCR is poor (garbled text, missing characters), Claira's output will reflect that. Garbage in, garbage out — switching to Image mode is often the right fix.
• Documents are analyzed individually. Claira does not cross-reference between documents. If the author is only named in a cover email but not in the attached report, the report will not inherit that author.
• Complex cases need human judgment. A document with five plausible dates requires a reviewer to decide which one is "primary." Claira will follow your prompt's instructions, but those instructions may not cover every edge case.

Need help? Contact support@claira.to