Best Practices

Tips, patterns, and recommendations for getting the most out of DocLD.

Document Preparation

File Quality

• Use high-resolution scans — For scanned PDFs and images, 300 DPI or higher improves OCR accuracy
• Prefer native digital documents — PDFs with embedded text layers parse faster and more accurately than scanned pages
• Avoid heavy compression — JPEG artifacts and low-quality scans can degrade text extraction

Format Choice

When OCR Helps

OCR runs automatically when:

• The file is an image (PNG, JPEG, TIFF, etc.)
• A PDF has no extractable text (scanned document)
• Parsing yields little or no text

For mixed documents (some pages scanned, some digital), DocLD handles each page appropriately.

Chunking and Retrieval

Chunking Strategy

DocLD uses layout-aware chunking. You can tune settings per knowledge base:

• Semantic boundaries — Splitting at paragraph and section boundaries preserves context
• Size limits — Default chunk sizes balance retrieval relevance and context length
• Overlap — Optional overlap between chunks helps queries that span boundaries

See the RAG Setup Guide for configuration details.

Structuring Knowledge Bases

• One domain per knowledge base — Keep related documents together (e.g., “HR Policies”, “Contract Library”)
• Curate content — Exclude irrelevant sections or documents that add noise
• Update regularly — Remove outdated content; add new documents as they become available

Extraction

Schema Design

• Write clear field descriptions — Tell the model where to find data: “The invoice number, usually at the top near the date”
• Use appropriate granularity — Prefer specific fields (e.g., invoice_date, due_date) over vague ones (dates)
• Add schema-level instructions — Explain edge cases: “If tax is shown separately, extract it as its own field”

Confidence Handling

Use the Custom Extraction Guide for schema examples and field types.

RAG and Chat

Query Optimization

• Be specific — “What is the vacation policy for full-time employees?” works better than “vacation?”
• Provide context — For multi-document bases, mention document type or topic when relevant
• Use follow-ups — Conversation context improves multi-turn answers

Citation Quality

• Check retrieval settings — Adjust top_k and threshold in RAG configuration
• Enable reranking — Improves relevance of retrieved chunks
• Choose response mode — Use thorough for complex questions needing more citations

Knowledge Base Organization

• Single topic per KB — Avoid mixing unrelated domains
• Appropriate size — Not too small (limited coverage) or too large (noise and slower retrieval)
• Related content — Documents should logically relate to each other

Performance and Cost

Batching

• Batch uploads — Use the batch API for multiple documents instead of sequential uploads
• Batch extraction — Run extractions across many documents in a single job
• Batch document addition — Add multiple documents to a knowledge base in one request

Reprocessing

• Monitor status — Use the dashboard or API to track processing status
• Reprocess failed documents — Fix issues (e.g., file corruption) and retry from the document detail page
• Update and reprocess — When chunking or vectorization settings change, reprocess affected documents

Monitoring

• Use Analytics — Track usage, quality metrics, and costs
• Review feedback — Thumbs up/down and user signals help tune retrieval and prompts
• Check confidence scores — Low scores may indicate schema or document quality issues

Quick Reference