nemo-retriever

name	nemo-retriever
description	Ingest and query document collections with NVIDIA NeMo Retriever-style retrieval workflows.

NVIDIA NeMo Retriever Workflows

Use this skill whenever the user asks to search, summarize, or extract facts from a corpus of uploaded documents, filings, PDFs, images, spreadsheets, presentations, audio, or video.

Ingestion pipeline

1. Identify the source material:
   • PDF widgets / uploaded PDFs → list_pdfs() then pdf_extract(name, page_range)
   • Images containing tables/charts → vision_qa.understand_image(...)
   • Audio / video clips → gemma_audio.transcribe_audio(...)
   • Web pages → fetch_url(url)
2. For large documents (>800 pages or many files), split ingestion into batches and summarize each batch before final synthesis.
3. If the material is in a non-English language and you need to embed or compare it, use translate(text, target_language="English") first.
4. Durable user memory is written automatically by the server's ingestion pipeline; recall_user_memory(query, k) is read-only and retrieves facts the user has accumulated across prior conversations.

Table extraction rule

When an OCR or PDF-extraction task is explicitly about tables, return the full table as an emit_table_artifact(columns=..., rows=..., name=..., description=...) — do not summarize the table into prose or drop rows. Include a concise description explaining the source document/page.

Use the dedicated nemotron_ocr.extract_text_from_image(..., return_table=True) tool for image-based tables; it understands layout and returns structured rows when possible.

Query workflow

1. Reformulate the user's question into 1-3 retrieval queries.
2. Search the extracted content with the appropriate tool:
   • PDFs → search_pdf(query, k=5)
   • Widget data tables → query_widget_data(sql=...)
   • Web → web_search(query, k=5)
3. Rerank by relevance: prefer primary sources, recent filings, and tables that directly answer the question.
4. Synthesize a concise answer and cite every claim with cite_source(text, source, source_url).

Output

• Factual summaries → emit_markdown_artifact(...)
• Extracted tables → emit_table_artifact(...)
• Do not return raw OCR dumps as chat text; always summarize.