NeMo Retriever Document Retrieval Skill
Search, extract, quote, and answer questions across document, image, audio, and video corpora.
The skill permits Bash, Write, and Read, and directs dependency installation, remote repository cloning, model downloads, remote endpoint use, and index writes. No malicious behavior is shown, but user confirmation, least-privilege scoping, sensitive-document handling, data-flow disclosure, rollback, and deletion guidance are incomplete, so points are deducted. NVIDIA provenance, Apache-2.0 licensing, and a security reporting path provide limited positive evidence.
The workflow, format coverage, failure modes, and recovery paths are fairly detailed, but default LanceDB filenames and table names conflict across SKILL.md, setup.md, ingest.md, and query.md. Some commands, scripts, and parameters are not sufficiently supported by the supplied files, and static review cannot verify execution, so points are deducted and the static ceiling applies.
Triggers, supported formats, and several non-fit scenarios are clearly stated for multi-file retrieval, extraction, transcription, and aggregation. However, the core workflow depends heavily on local GPU capability, model downloads, external services, and host packages; Chinese-language behavior and mainland-China network reachability are not adequately addressed, and parts of the workflow remain PDF-centric, so points are deducted.
The skill uses progressive reference documents for installation, setup, querying, CLI details, and troubleshooting, and declares Apache-2.0. It lacks skill-level authorship, versioning policy, changelog, explicit maintenance responsibility, a fuller FAQ, and pinned dependency governance; BENCHMARK also records missing recommended schema fields, so points are deducted.
The skill covers multimodal indexing and retrieval and specifies source/page-location fields that could support directly usable answers. However, no representative outputs are statically verifiable, benchmark effectiveness is only 45%/47%, and known limitations affect recall, citation indexing, and image transcription, so only limited credit is warranted.
evals.json and BENCHMARK.md provide a dated evaluation summary, task composition, and metrics, giving limited auditability. There are no raw results, committed test suites, key-path CI workflows, or independent reproduction artifacts, so the static evidence is insufficient for a higher score.
- Default index filenames and table names conflict across the documentation, which may cause queries to miss an index that was created successfully.
- Installation may clone remote code, download models, and install system packages; handling of sensitive documents and data-flow risks for remote embedding/VLM endpoints are not adequately documented.
- The benchmark reports only 45%/47% effectiveness and provides no raw results or reproducible test evidence; its PASS verdict should not be treated as independent validation.
- Audio/video, Office, OCR, and remote-model features may require extra software, GPUs, or overseas network services; mainland-China users should verify reachability and alternatives first.
What does this skill do, and when should you use it?
This skill instructs an AI agent to use the NeMo Retriever Library for multi-file and non-PDF corpora. The retriever CLI ingests a folder into LanceDB and provides vector search for later queries. It supports PDFs, scanned images, Office files, HTML, TXT, audio, and video, along with page filtering, verbatim citations, corpus-level aggregation, and chart or image-caption retrieval. It is not intended for file editing, web browsing, single-file plain-text lookups, or fine-tuning.
Guides installation of the NeMo Retriever Library when the retriever command is unavailable; runs retriever ingest to index a folder in LanceDB; runs retriever query for corpus search and question answering; handles PDF, JPG, PNG, TIFF, DOCX, PPTX, HTML, TXT, MP3, WAV, M4A, MP4, and MOV inputs; and supports page filtering, verbatim quotations with citations, corpus-level aggregation, and chart or image-caption hits.
- A researcher needs answers synthesized from a folder of PDF reports.
- An operations user needs information extracted from scanned forms or image files.
- A knowledge worker needs to search across DOCX, PPTX, HTML, or TXT documents.
- A media analyst needs to search, transcribe, or extract information from audio or video corpora.
- A user needs corpus-wide filtering, cited quotations, or aggregate results rather than a simple keyword lookup.
What are this skill's strengths and limitations?
- Supports documents, scanned images, audio, and video.
- Designed for multi-file and non-PDF corpora.
- Covers filtering, verbatim citation, aggregation, and other non-semantic operations.
- Uses the dedicated retriever CLI with LanceDB indexing.
- The supplied material does not provide complete installation, setup, or query parameter examples.
- Requires the NeMo Retriever Library, retriever CLI, and LanceDB environment.
- Explicitly excludes file editing, web browsing, single-file plain-text lookups, and fine-tuning.
- Mixed-format folders require extension inventory because auto mode can silently drop unsupported file types.
How do you install this skill?
Install the skill through the repository’s documented flow: npx skills add nvidia/skills --skill nemo-retriever --yes. If retriever is missing, follow references/install.md to install the NeMo Retriever Library; the supplied source does not include the complete installation command.
How do you use this skill?
After installing it in an Agent Skills-compatible client, give a concrete request such as: “Search this folder of PDFs for every contract termination clause and summarize the results with page citations.” The first use requires indexing; subsequent questions use retriever query. Exact arguments are documented in the skill’s setup, query, and CLI references.
How does this skill compare with similar options?
The skill explicitly prefers retriever over native Read or Grep for multi-file or non-PDF corpora, while excluding single-file plain-text lookups.