Data Sampler Plugin — Instructions

Overview

Two tools for preparing text for LLM consumption:

• sample — Split long texts into optimized chunks using 3 strategies
• load_dataset — Load sample rows from HuggingFace datasets via the Datasets Server API

---

sample Tool

Parameters

Parameter	Type	Default	Description
text	string (req)	—	Input text to chunk
strategy	string	"priority"	"priority" (keyword frequency + position), "relevance" (BM25 vs query), "length" (fixed-size)
chunkSize	number	2000	Target chars per chunk
overlap	number	0	Overlapping chars between chunks (length only)
maxChunks	number	unlimited	Max chunks to return
query	string	—	Required for "relevance" strategy

Output

Use Cases

• priority: Extract the most information-dense segments from long articles
• relevance: Find segments matching a specific query (e.g. "budget figures")
• length: Split into uniform segments for parallel processing

---

load_dataset Tool

Parameters

Parameter	Type	Default	Description
dataset	string (req)	—	HF dataset name (e.g. "squad", "imdb", "ibm/duorc")
split	string	"train"	Dataset split ("train", "test", "validation")
config	string	auto-detect	Dataset config/subset (required if multiple configs exist)
max_samples	number	5	Number of rows to fetch (max 100)
query	string	—	Optional BM25 search query to filter rows
format	string	"text"	"text" (key: value lines), "json" (raw JSON), "prompt" (=== Sample N ===), "fewshot" (Input:/Output: pairs)
template

Output

API Used

The tool uses the HuggingFace Datasets Server (datasets-server.huggingface.co) REST API — no extra npm dependencies. Public datasets are accessible without authentication.

Use Cases

• Load few-shot examples from SQuAD, IMDB, or custom datasets
• Inject dataset samples into prompts for in-context learning
• Search across dataset rows with BM25 full-text search
• Pick specific rows by index (indices) for curated example sets
• Format samples as few-shot Input:/Output: pairs ready to paste into a prompt (format: "fewshot")

Few-Shot Prompt Injection (fewshot format)

format: "fewshot" renders each row as labeled Input Column: value / Output Column: value lines — ready to be pasted into a system or user prompt as in-context examples:

• input_columns selects which fields appear as inputs (default: all except the output column)
• output_column selects the answer field (default: the last column in the row)
• Combine with indices: "3,17,42" to curate exactly which examples enter the prompt

---

Tips

• The load_dataset tool auto-discovers the dataset config if not specified. If multiple configs exist, it asks you to specify one.
• The sample tool's "relevance" strategy pairs well with load_dataset — load samples, then chunk the context field.
• For rate limiting: the HF Datasets Server has a max of 100 rows per call. Make multiple calls for more samples.