---

name: generation-and-refinement description: "Use FlashRAG generators, prompt templates, OpenAI/vLLM/HF/FastChat backends, multimodal generation, refiners/compressors, and judgers as reusable components." disable-model-invocation: true metadata: disco-role: operating license: MIT

Generation and Refinement

Use this sub-skill when a task asks you to configure, instantiate, call, or debug FlashRAG generation-side components rather than build indexes, assemble full RAG methods, or report final metrics.

Route by Task

• Text generation: Use flashrag.utils.get_generator(config) and generator.generate(...); see generator-api.md.
• Prompt construction: Use flashrag.prompt.PromptTemplate to convert questions, retrieval results, messages, and previous generations into backend-ready strings or chat messages; see generator-api.md.
• Backend selection: Choose framework: hf, vllm, fschat, or openai, then align model paths, credentials, optional packages, and generation_params; see model-and-backend-options.md.
• Multimodal generation: Use HF-backed multimodal generators with message blocks containing text and image content; see model-and-backend-options.md.
• Refiners and compressors: Use flashrag.utils.get_refiner(config) or a concrete refiner to compress retrieved documents or prompts; see refiner-and-judger.md.
• Judgers: Use flashrag.utils.get_judger(config) when a pipeline needs a component that decides whether or how strongly to retrieve; see refiner-and-judger.md.
• Preflight checks: Run the bundled safe checker before loading large models or calling APIs: python skills/flashrag/sub-skills/generation-and-refinement/scripts/inspect_generation_config.py path/to/config.yaml; see troubleshooting.md.

Boundaries

This sub-skill owns generator, prompt, refiner, compressor, and judger components. For retrieval model setup, index files, corpus processing, and reranking, use the retrieval/indexing sub-skill. For complete methods such as Self-RAG, FLARE, Adaptive-RAG, Trace, or end-to-end experiment scripts, use the pipelines/methods sub-skill. For evaluation metrics or WebUI behavior, use the evaluation/WebUI sub-skill.

Safety Notes

• Do not paste real OpenAI or Azure credentials into examples, reports, or generated configs; use placeholders or environment variables.
• Treat generator_model_path, refiner_model_path, and judger model paths as user-supplied model identifiers or local paths; validate intent before downloading or loading large models.
• The bundled inspector is static and does not import FlashRAG, load models, download tokenizers, or make API calls.