Grounded Citations - Setup Guide¶
Source: nousresearch/hermes-agent (22 installs) Category: Research / Quality Assurance License: MIT · Platforms: Linux, macOS, Windows Dependencies: Hermes Agent, web retrieval tools (web_search/web_extract)
Every claim from an outside source gets an inline numbered citation and a Sources: reference list, Perplexity-style. A ledger script owns the url → [n] mapping so numbers and URLs come from retrieval, never from model memory. For high-stakes work, the same ledger doubles as a fact-checking chain - verbatim quotes are verified against source text, and model-knowledge claims are flagged [unverified].
What It Does¶
| Capability | How |
|---|---|
| Inline citations | Every external claim gets [n] linking to numbered source |
| Source ledger | url → [n] mapping owned by retrieval, not model memory |
| Verbatim verification | Quotes rejected unless they literally appear in fetched text |
| Unverified flagging | Model-knowledge claims tagged [unverified] |
| Evidence validation | verify --evidence fails drafts with unsupported claims |
| Multi-format support | Works in chat, markdown, PDF, docx, and slides |
How It Works¶
┌──────────────┐ ┌──────────────┐ ┌──────────────┐
│ Research │────▶│ Source │────▶│ Output │
│ Question │ │ Ledger │ │ with [n] │
│ │ │ url → [1..N] │ │ citations │
└──────────────┘ └──────┬───────┘ └──────┬───────┘
│ │
┌────────▼───────┐ ┌────────▼───────┐
│ Fetch pages │ │ verify │
│ Extract text │ │ --evidence │
└────────────────┘ └────────────────┘
Installation¶
Via skills.sh (Recommended)¶
npx skills add https://github.com/nousresearch/hermes-agent --skill grounded-citations
Direct from Hermes Agent Repo¶
git clone --depth 1 https://github.com/nousresearch/hermes-agent.git /tmp/hermes-agent
cp -r /tmp/hermes-agent/skills/research/grounded-citations ~/.hermes/skills/
Basic Usage¶
Research with citations¶
> Load grounded-citations skill
> Research: What's the current state of RAG retrieval techniques?
> Use grounded citations
Hermes will search, fetch pages, extract text, and produce:
Recent advances in RAG retrieval have shifted toward agentic
multi-step retrieval [1] and hybrid dense-sparse approaches [2].
ColBERT-style late interaction has shown 15-20% improvement over
single-vector methods on the BEIR benchmark [3].
Sources:
[1] https://arxiv.org/abs/2401.xxxxx - "Agentic RAG: A Survey"
[2] https://blog.langchain.dev/... - Hybrid Search in Production
[3] https://arxiv.org/abs/2004.12832 - ColBERT (original paper)
Verify existing content¶
> Verify the citations in /home/user/report.md --evidence
This checks every [n] citation - if a claim doesn't trace to source text, it fails.
Citation Format Examples¶
In chat:
...as demonstrated by Smith et al. [1].
In markdown:
...as demonstrated by Smith et al. [^1].
[^1]: https://example.com/paper - "Title Here" (2024)
In documents (PDF/docx): Superscript numbers with endnote-style source list.
When to Use vs When Not¶
| ✅ Use For | ❌ Don't Use For |
|---|---|
| Research reports & briefings | Casual conversation |
| Factual claims about the world | Personal opinions/advice |
| Content for publication | Creative writing |
| Stakeholder-facing documents | When speed matters more than accuracy |
| Academic or professional writing | Model self-knowledge about its own capabilities |
Integration with Other Skills¶
| Skill | How It Integrates |
|---|---|
| research-paper-writing | Feeds citation chain into academic paper pipeline |
| arxiv | Auto-cites arXiv papers with proper BibTeX-ready references |
| web-search-duckduckgo | Primary retrieval source for web citations |
| blogwatcher | Citations from blog monitoring feeds |
Tips¶
- Trust the ledger, not memory: The
url → [n]mapping is deterministic - model never invents URLs - Verbatim quotes only: If the text isn't literally in the fetched page, the citation fails
[unverified]is honest: Better to flag model knowledge than fabricate a source- Re-run
verify --evidenceafter edits: Adding/removing claims may break citation chains - Use with
research-paper-writingfor academic work: This skill handles web citations; that skill adds BibTeX and formatting
Troubleshooting¶
| Problem | Likely Cause | Fix |
|---|---|---|
All citations flagged [unverified] |
Web retrieval tools down | Check web_search/web_extract availability |
| "Evidence verification failed" | Source page changed since fetch | Re-fetch and re-verify |
| Missing citations in output | Skill not loaded before research | Load skill before starting research |
| Too many citations | Model over-citing | Narrow scope or use --citation-density low |
← Skills Catalog | nousresearch/hermes-agent on skills.sh
Powered by CorpusIQ
Integration with CorpusIQ¶
CorpusIQ is the connector layer that lets AI assistants read the systems a business already runs on - accounting, payments, CRM, email, files and analytics - and answer with the source named. It is read-only by design: it retrieves data and never writes to, or acts in, a connected system, which is what makes it safe to point at live accounts.
This page is part of the CorpusIQ directory, the reference we maintain for the MCP server and agent-skill ecosystem. Run alongside CorpusIQ, a third-party server like this one can be combined with a company's own business data in a single answer.