Usage Examples

These examples show practical settings for different document collections. Paths are examples; replace them with local paths on your machine.

Technical Documentation Library

Use this for API references, Markdown docs, JSON/YAML config examples, and generated documentation.

Documents Directory: ~/Documents/tech-library
Vector Store Directory: ~/.lmstudio/tech-library-db
Chunk Size: 768
Chunk Overlap: 150
Chunking Strategy: sentence
Min Chunk Length: 20
Retrieval Limit: 7
Retrieval Affinity Threshold: 0.55
Max Concurrent Files: 3
Enable OCR: false

Example queries:

• "How do I connect to PostgreSQL using Python?"
• "What are the differences between var, let, and const in JavaScript?"
• "Find examples of Kubernetes YAML resource limits."

Notes:

• Sentence chunking works well for explanatory docs.
• JSON, JSONL, YAML, YML, CSV, TSV, RST, and LOG files are indexed as text.
• Increase chunk size for long API reference sections.

Research Paper Collection

Use this for PDF-heavy academic collections.

Example queries:

• "What approaches to attention mechanisms are discussed?"
• "Compare reinforcement learning algorithms for robotics."
• "Which datasets are used for object detection?"

Notes:

• Keep OCR off for text-based PDFs.
• Enable OCR only for scanned papers.
• Use a higher retrieval limit when you want cross-paper comparison.

Office Document Archive

Use this for DOCX and ODT collections such as policies, contracts, procedures, and reports.

Example queries:

• "What is the refund policy?"
• "Find non-compete language in contracts."
• "Summarize the onboarding procedure."

Notes:

• DOCX parsing uses mammoth.
• ODT parsing uses adm-zip when available and falls back to best-effort XML text extraction.
• Sentence chunking usually improves retrieval for prose documents.

Legal Document Archive

Use this for contracts, regulations, case files, and scanned legal records.

Example queries:

• "Find all contracts with non-compete clauses."
• "What are the standard terms for intellectual property rights?"
• "Show precedents for breach of contract cases."

Notes:

• Higher thresholds reduce low-confidence matches.
• Lower concurrency is safer for OCR-heavy collections.
• Review citations manually for legal work.

Personal Knowledge Base

Use this for notes, articles, saved web pages, recipes, and ebooks.

Example queries:

• "What did I save about productivity techniques?"
• "Find the chocolate cake recipe."
• "What notes mention quarterly planning?"

Notes:

• Character chunking is fast and works well for mixed content.
• Raise minChunkLength if short notes or page fragments create noisy retrieval results.
• Use exclude patterns for folders you do not want indexed.

Log and Data File Collection

Use this for operational logs, CSV exports, JSONL traces, and config directories.

Example queries:

• "Find errors related to failed authentication."
• "Which config files mention the staging endpoint?"
• "Show rows related to customer churn."

Notes:

• Character chunking is usually better for logs and structured text because sentence boundaries may not exist.
• Use a lower threshold for varied log wording.
• Consider exclude patterns for generated or oversized noisy files.

Long User Queries

If users paste large text blocks into chat, set Max Query Length to match the embedding model limit. The plugin truncates only the query used for embedding; the original user prompt is still preserved in the final prompt template.

Recommended starting values:

• 512 tokens for smaller embedding models.
• 1024 tokens for larger local embedding models.
• 2048 tokens only if the embedding model supports long inputs reliably.

General Tuning Rules

• Start with a small subset and confirm retrieval quality.
• Use sentence chunking for prose and character chunking for logs, tables, and configs.
• Keep maxConcurrentFiles low for PDFs and OCR.
• Raise minChunkLength to remove headers, page numbers, and boilerplate.
• Reindex after changing embedding models or chunking settings.