DataFlow-Skills

DataFlow 开发专家上下文加载器。当用户在 DataFlow 仓库中进行开发时触发， 涵盖：新建算子/Pipeline/Prompt、诊断报错、规范审查、 以及感知仓库变更并建议更新知识库。 Trigger: user is developing in DataFlow repo, asks to create operator/pipeline/prompt, encounters errors, wants code review, or asks about operators.

Reasoning-guided pipeline planner that generates standard DataFlow pipeline code

Reference documentation for the BenchDatasetEvaluatorQuestion operator. Extended version of BenchDatasetEvaluator with question and subquestion support. Use when: evaluating answers with question context or multiple subquestions.

Reference documentation for the BenchDatasetEvaluator operator. Covers the constructor, two comparison modes (match/semantic), and pipeline usage. Use when: comparing predicted answers against ground truth answers in benchmark evaluation.

Reference documentation for the PromptedEvaluator operator. Use when: scoring text quality with LLM without filtering rows.

Reference documentation for the Text2QASampleEvaluator operator. Use when: evaluating QA pair quality across multiple dimensions.

Reference documentation for the UnifiedBenchDatasetEvaluator operator. Use when: evaluating model answers on benchmark datasets.

Reference documentation for the GeneralFilter operator. Covers the constructor, rule-based filtering logic, and pipeline usage notes. Use when: filtering rows based on column value conditions that can be expressed as lambda functions without LLM calls.

Reference documentation for the KCenterGreedyFilter operator. Covers the constructor, K-Center Greedy algorithm behavior, embedding serving requirements, and pipeline usage notes. Use when: downsampling a large dataset by semantic diversity using embedding vectors.

Reference documentation for the PromptedFilter operator. Covers the constructor, actual scoring and filtering behavior, and pipeline usage notes. Use when: filtering rows based on LLM semantic quality judgment rather than simple rule-based conditions.

Reference documentation for the PandasOperator operator. Use when: applying custom DataFrame transformations without LLM.

Reference documentation for the PromptedRefiner operator. Use when: refining text with LLM, overwriting original column.

Reference documentation for the BenchAnswerGenerator operator. Covers the constructor, full run() signature, actual generation behavior, and integration notes for unified bench evaluation pipelines. Use when: generating model answers from benchmark question rows before passing the dataframe into UnifiedBenchDatasetEvaluator.

Reference documentation for the ChunkedPromptedGenerator operator. Covers the constructor, file-path based chunking flow, actual prompt construction, and output file writing behavior. Use when: the dataframe stores file paths, the file content may exceed a single LLM context window, and you want the generated results written into new text files.

Reference documentation for the EmbeddingGenerator operator. Covers the constructor, embedding serving requirements, actual dataframe flow, and runnable pipeline usage. Use when: converting one text column in a dataframe into embedding vectors for retrieval, clustering, similarity search, or downstream vector-based operators.

Reference documentation for the FormatStrPromptedGenerator operator. Covers the constructor, prompt template restrictions, placeholder-to-column mapping, actual prompt-building logic, and runnable example usage. Use when: one generation task needs multiple dataframe columns combined into a single prompt through a template.

Reference documentation for the PromptedGenerator operator. Covers constructor parameters, run() signature, actual row-processing behavior, and pipeline usage notes. Use when: integrating PromptedGenerator into a DataFlow pipeline for single-field LLM generation.

Reference documentation for the RandomDomainKnowledgeRowGenerator operator. [Purpose] Calls an LLM repeatedly with the same domain-generation prompt and writes the generated results into one column of an existing DataFrame. [When to use] Use it when you already have a seed DataFrame with the exact target row count and want to fill one output column with domain-specific generated content. If you need to build prompts from existing row fields, use PromptedGenerator or FormatStrPromptedGenerator instead.

Reference documentation for the RetrievalGenerator operator. [Purpose] Reads one text column from storage, forwards every non-empty row to `llm_serving.generate_from_input(...)`, and writes the returned list into a new output column. [Default backend] Use `LightRAGServing` by default. [Important] `run()` is async. The operator itself does not initialize the serving object; it only awaits `llm_serving.generate_from_input(llm_inputs, system_prompt)`.

Reference documentation for the Text2MultiHopQAGenerator operator. [Purpose] Generates multi-hop QA pairs from one text column and writes two output columns: one for `qa_pairs` and one for metadata. [When to use] Use it when you want reasoning-style QA pairs derived from longer text chunks. If only simple single-hop QA is needed, use `Text2QAGenerator` instead.