papi

Files (2)

• SKILL.md
• commands.md · references/

SKILL.md (first lines)

# Paper Reference Assistant (CLI)

**Entry point skill.** Use `papi` CLI first; MCP RAG tools only when CLI is insufficient.

For specialized workflows, invoke dedicated skills:
- `/papi-ask` — RAG queries requiring synthesis
- `/papi-verify` — verify code against paper
- `/papi-compare` — compare papers for decision
- `/papi-ground` — ground responses with citations
- `/papi-curate` — create project notes

## Setup

```bash
papi path   # DB location (default ~/.paperpipe/; override via PAPER_DB_PATH)
papi list   # available papers; titles are truncated, never cut the output with head
papi search "2411.16820"   # is a paper already in the DB? by arXiv ID or name
```

A paper is in the database when the top `papi search` hit for its arXiv ID or name is that paper (an ID hit
shows `Matches: arxiv_id`). Check this way before saying a paper is missing; a topic grep of `papi list`
misses papers whose titles don't contain your words.

## When NOT to Use MCP RAG

- Paper name known → `papi show <paper> -l summary`
- Exact term search → `papi search --rg "term"`
- Checking equations → `papi show <paper> -l eq`
- Only use RAG when above methods fail or semantic matching required

## Decision Tree

| Question | Tool |
|----------|------|
| "What does paper X say about Y?" | `papi show X -l summary`, then `papi search --rg "Y"` |
| "Does my code match the paper?" | `/papi-verify` skill |
| "Which paper mentions X?" | `papi search --rg "X"` first, then `leann_search()` if no hits |
| "Compare approaches across papers" | `/papi-compare` skill or `papi ask` |
| "Need citable quote with page number" | `retrieve_chunks()` (PQA MCP) |
| "Cross-paper synthesis" | `papi ask "..."` |

## Search Commands

```bash
papi search --rg "query"                        # literal text match (fast, no LLM)
papi search --rg --regex "term one|term two"    # regex or OR match; --regex is required
papi search "query"                   # ranked BM25
papi search --hybrid "query"          # ranked + exact boost
papi search "query" -p paper1,paper2  # limit search to specific papers
papi ask "question"                   # PaperQA2 RAG
papi ask "question" --backend leann   # LEANN RAG
papi notes {name}                     # open/print implementation notes
```

## Search Escalation (cheapest first)

1. `papi search --rg "X"` — exact text, fast, no LLM
2. `papi search "X"` — ranked BM25 (requires `papi index --backend search` first)
3. `papi search --hybrid "X"` — ranked + exact boost
4. `leann_search()` — semantic search, returns file paths for follow-up
5. `retrieve_chunks()` — formal citations (DOI, page numbers)
6. `papi ask "..."` — full RAG synthesis

## MCP Tool Selection (when papi CLI insufficient)

| Tool | Speed | Output | Best For |
|------|-------|--------|----------|
| `leann_search(index_name, query, top_k)` | Fast | Snippets + file paths | Exploration, finding which paper to dig into |
| `retrieve_chunks(query, index_name, k)` | Slower | Chunks + formal citations | Verification, citing specific claims |
| `papi ask "..."` | Slowest | Synthesized answer | Cross-paper questions, "what does literature say" |

- Check available indexes: `leann_list()` or `list_pqa_indexes()`
- Embedding priority: Voyage AI → Google/Gemini → OpenAI → Ollama

## Adding Papers

```bash
papi add 2303.13476                   # arXiv ID
papi add https://arxiv.org/abs/...    # URL
papi add 2303.13476 1706.03762 "Attention Is All You Need"  # multiple at once (mixed sources OK)
papi add --pdf /path/to.pdf           # local PDF
papi add --pdf "https://..."          # PDF from URL
papi add --from-file papers.bib       # bulk import
```

## Per-Paper Files

Located at `{db}/papers/{name}/`: `equations.md`, `summary.md`, `source.tex`, `notes.md`, `paper.pdf`, `figures/`.

If agent can't read `~/.paperpipe/`, export to repo: `papi export <papers...> --level equations --to ./paper-context/`
Use `--figures` to include extracted figures in export.

See `references/commands.md` for full command reference and per-file details.

From hummat/paperpipe, MIT.

More from hummat/paperpipe