paper-research-agent 0.5.1

Local-first multi-agent CLI for paper understanding, critical debate, literature intelligence, Scribe-powered writing, and figure generation.

Research Agent

Local-first conversational CLI for research paper understanding, critical discussion, and literature intelligence. Multi-agent (Analyst + Critic + Searcher + MemoryKeeper) with explicit slash commands and LLM tool calling — pick whichever feels natural per turn.

Requirements

• Python 3.11+
• An OpenRouter API key (any OpenAI-compatible endpoint works; OpenRouter is the default)

Install

Distribution name on PyPI is paper-research-agent (the Python import name stays research_agent). Currently published to Test PyPI while the release stabilises; a production PyPI upload follows once the Test PyPI build has soaked.

Option A — From Test PyPI (current)

The actual runtime dependencies (PyMuPDF, ChromaDB, openai, …) only live on real PyPI, so you need both indexes:

pip install \
  --index-url https://test.pypi.org/simple/ \
  --extra-index-url https://pypi.org/simple/ \
  paper-research-agent

Or with pipx for an isolated install:

pipx install \
  --index-url https://test.pypi.org/simple/ \
  --pip-args="--extra-index-url https://pypi.org/simple/" \
  paper-research-agent

Option B — From PyPI (once published)

pip install paper-research-agent        # or: pipx install paper-research-agent

Option C — From source (development)

git clone https://github.com/ynsun-tw/research-helper.git
cd research-helper
pip install -e ".[dev]"

After any of the above, research --help should list the writing suite (write, review, check, style) alongside config and insights, and the conversational REPL is one research away.

Latest published build: 0.5.0 (M5 polish — figure generation, doctor, performance, CI, packaging) on Test PyPI — project page.

Configure

# API key from https://openrouter.ai/keys
research config set api_key <your-openrouter-key>

# Default model: deepseek/deepseek-chat — change to any OpenRouter model id
research config set model anthropic/claude-3.5-sonnet
research config set language zh           # or: en (default), zh, 中文

research config show

Configuration is stored at ~/.research-agent/config.yaml (file mode 600).

Key	Default	Description
api_key	—	OpenRouter API key
model	deepseek/deepseek-chat	Model slug on OpenRouter
base_url	https://openrouter.ai/api/v1	API base (change only if self-hosting a proxy)
app_title	Research Agent	Sent as X-Title header to OpenRouter
app_url	https://github.com/research-agent	Sent as HTTP-Referer header
language	en	Agent reply language: en or zh
alert_threshold	0.8	Cosine similarity threshold for the parked-idea alert that fires on /read (range [0.0, 1.0]; lower = more reminders)

All local state lives under ~/.research-agent/: memory.db (SQLite), chroma/ (vector indexes for ideas + discussions), and cache/ (downloaded PDFs).

Usage

Running research with no arguments drops you into the conversational REPL. Everything else happens inside it.

research                                       # enter the REPL
research config set api_key sk-or-...          # the only remaining subcommand
research --help

Inside the REPL you can either type slash commands for explicit control or plain text to let the LLM pick the right tool.

Slash commands

Command	What it does
/search [--mode theoretical|applied|group:<author>] <keywords>	Paper search with LLM relevance scoring; primary source is arXiv with a Semantic Scholar fallback if arXiv rate-limits or errors. Sorted by score; flags papers already in your library. Optional --mode biases the candidate set: theoretical (analysis / proofs), applied (benchmarks / experiments), or group:"<author name>" (quote multi-word names)
/history [N]	Recent /search queries across sessions, with hit counts and read markers
/recall <query>	Semantic search across past REPL discussions (cross-session)
/read <arxiv-id | title | path.pdf>	Download + Analyst + Critic; sets the conversation anchor; auto-marks the queue entry done if present
/discuss <idea or follow-up>	First turn: structured Analyst (contributions / impact / related work) + Critic (objections / score / suggestions). Follow-up turns: grounded prose, no re-scoring. Must run /read first.
/queue	List pending entries (alias for /queue list)
/queue add <id> [title…]	Save a paper for later (pending)
/queue list [all|pending|done|skipped|in_progress]	Filter the queue
/queue next	Preview the next pending entry without state change
/queue read	Load + analyze the next pending entry, auto-mark done
/queue done|skip|remove <id>	Manual state transitions
/cites [arxiv-id]	Papers that cite the anchor (or given) paper — forward references via Semantic Scholar
/refs [arxiv-id]	Papers cited by the anchor (or given) paper — backward references via Semantic Scholar
/refine	Ask Searcher to propose the next search query from your recent discussion (query + optional --mode + reason + confidence); interactively accept / edit / skip
/insights [--since 30d|7d|6m|all]	Deterministic Markdown summary of your local activity: papers (by year, top tags / authors / venues), ideas (by status, average critic score, most-debated, top-scored) and discussion volume. No LLM call - safe to run anywhere
/paper	Summary of the current anchor paper
/idea save [title]	Persist the active debate as a saved idea
/ideas	List saved ideas with their latest critic score
/idea show <id-prefix>	Show one idea + its full score history
/ideas update <id> [--status <s>] [--feedback <note>] [--condition "<phrase>"] [--clear-conditions]	Update status, log score feedback, or pin / clear activation conditions (multiple --condition flags allowed)
/help	List every slash command
/exit	Persist + flush vector indexes + quit

After every /read, Research Agent quietly checks your shelved / waiting ideas; if the paper looks topically related (cosine similarity ≥ alert_threshold, default 0.8), it prints a one-line banner with /idea show <prefix> shortcuts so you can revisit context you parked earlier. Tune the trigger via research config set alert_threshold 0.85 (range [0.0, 1.0]; lower = more reminders, higher = fewer false positives).

You can also pin activation conditions on a shelved idea — free-form phrases that describe what would unblock it (a dataset release, a checkpoint, a baseline result). Set them via /ideas update <id> --condition "FineWeb-Edu dataset" (repeatable in one command, clear with --clear-conditions). Every /search then scans incoming hits for those phrases (case-insensitive substring) and prints a "Shelved idea(s) may have an unblock" banner whenever a new paper mentions one — letting search results pull an idea back into your attention automatically.

Natural language → tools

Plain text is sent to the LLM, which has function-calling access to the backend. Available tools:

search_arxiv, recent_searches, recall_history, load_paper, discuss_idea, save_current_idea, list_ideas, queue_add, queue_list, queue_next, get_citations, get_references, suggest_search_refinement, research_insights.

The model is instructed to chain them: "open the BERT paper I searched last week" → recent_searches → load_paper. "read the next one on my list" → queue_next → load_paper. "what did we conclude about positional encodings?" → recall_history then a synthesized recap. "who built on this paper?" → get_citations on the anchor paper. "what does this paper rely on?" → get_references. "what should I search next?" → suggest_search_refinement → search_arxiv. "how am I doing this month?" → research_insights.

Quick start (5 minutes)

# 1. Install (pick one)
pipx install --index-url https://test.pypi.org/simple/ \
             --pip-args="--extra-index-url https://pypi.org/simple/" \
             paper-research-agent
#  or  →  pip install -e ".[dev]" from the repo root for a dev install

# 2. Verify (no API key needed yet)
research --version            # → research-agent 0.5.0
research doctor               # → environment health check (config, DB, disk, chromadb)

# 3. Configure
research config set api_key sk-or-...     # OpenRouter key from https://openrouter.ai/keys
research config set language zh           # or en (default)

# 4. Drive the REPL
research                                  # enter the conversational shell
# inside the REPL:
›  /search efficient transformer long context
›  /read 1706.03762
›  /discuss replace dense attention with top-k sparse attention
›  /idea save sparse-attention
›  /exit

# 5. Author with Scribe
research style train arxiv:2305.14314 arxiv:2301.07041   # learn your voice
research style fingerprint                                # build the fingerprint
research write introduction --context "sparse top-k attention" --output intro.md
research review intro.md --section introduction --interactive

# 6. Diagrams
research figure --type architecture --desc "three-layer sparse encoder"
research figure --type result --data "ours 85, baseline 80" --verify

# 7. Sanity checks
research check intro.md       # self-plagiarism scan against your training corpus
research insights --since 30d # Markdown rollup of recent activity

If anything looks off, research doctor prints a single Rich table with every check, its status, and a one-line hint. Set RESEARCH_AGENT_DEBUG=1 to see the full Python traceback when an unexpected error fires (otherwise you only get one coloured line).

› /search --mode applied efficient transformer long context
› /queue add 1706.03762 Attention Is All You Need
› /read 1706.03762
› /discuss replace dense attention with top-k sparse attention for 32k contexts
› /idea save sparse-routing-attention
› /ideas update <id-prefix> --status shelved --condition "FlashAttention-3 release"
› /refine                                    # ask Searcher for the next query
› /insights --since 30d                      # weekly research review
› /exit

See examples/end-to-end-demo.md for a full scripted walkthrough that exercises every feature (search → relevance scoring → queue → read → citation graph → two-phase debate → parked-idea alerts → activation conditions → dynamic refinement → cross-session recall → research insights) on a real paper.

Writing assistant (M4, in progress)

Train the upcoming Scribe agent on your own published papers so it writes in a voice that actually sounds like yours. Today the M4 surface covers sample import (S4.1.1); fingerprint analysis + draft generation + writing-review pipeline land in subsequent stories.

# Pull paragraphs from a folder of PDFs
research style train --dir ~/papers

# Or hand-pick sources (arXiv ids and local PDFs may be mixed)
research style train arxiv:2301.07041 ~/papers/my-thesis.pdf

# Inspect the corpus
research style show

style train parses each source, splits it into paragraphs, drops non-prose (references, acknowledgements, formula-dense methodology, single-sentence captions), and writes the survivors into the style_samples table under ~/.research-agent/memory.db. Re-running the same source replaces its prior samples by default; pass --append to accumulate instead.

Once samples exist, build a fingerprint that captures how you write:

research style fingerprint

The fingerprint is computed entirely offline (no LLM call) and lands at ~/.research-agent/style/fingerprint.json. It has three layers:

Layer	What it captures
Macro	abstract opener, intro opener, related-work organization (chronological / thematic / comparison), avg sections per paper
Micro	sentence-length distribution (avg / median / p10 / p90), avg paragraph length, top transition words & their per-100-sentence rates, hedging / confidence / passive rates, type-token ratio
Markers	dominant citation format (latex_cite / bracket_num / author_year / mixed), figure & table reference style (Figure vs Fig.), em-dash usage, your top section titles

research style show prints the corpus inventory and (if present) the fingerprint summary, side by side.

Draft a section with Scribe

research write abstract --words 250
research write introduction \
  --context "this paper studies sparse top-k attention for 32k contexts" \
  --versions 3 \
  --output drafts/intro.md

research write <section> invokes the Scribe agent, which produces three stylistic variants by default — concise, technical depth, and narrative arc — by issuing the LLM calls in parallel (use --sequential to disable). Each draft is rendered in its own Rich panel with a one-line note explaining how it differs from the others; pass --output drafts.md to also save the bouquet to disk.

Supported section names (aliases in parentheses): abstract, introduction (intro), related_work (related), method (methods / approach), results (experiments / evaluation), discussion, conclusion.

When ~/.research-agent/style/fingerprint.json exists, the Scribe mimics it (sentence length, transitions, hedging vs confidence balance, citation format). Without a fingerprint it falls back to neutral academic prose and says so in each draft's style note.

Context-aware writing

When you pass --context "<description>", the Scribe doesn't just parrot the description — it also pulls related material from your memory store and injects it into the prompt:

• Related ideas from your library (semantic similarity ≥ 0.5, top 3) — title, status, last critic score, summary.
• Recent cross-session discussion excerpts that look topically relevant (top 3).
• Existing drafts you point it at with --check-against PATH (repeatable) — body included verbatim (truncated to ~2 KB each) with an explicit instruction not to duplicate or contradict.

research write conclusion \
  --context "sparse top-k attention for 32k contexts" \
  --check-against drafts/intro.md \
  --check-against drafts/method.md \
  --versions 3

The CLI prints a one-line Scribe context: user context, 2 related idea(s), 1 discussion excerpt(s), 2 draft(s) to stay consistent with summary before the panels so you can see what the agent saw.

Review a draft (Analyst + Critic → Scribe)

research review drafts/intro.md --section introduction --output drafts/intro.review.md

research review <file> runs the auto-review pipeline:

1. Analyst (writing mode) — flags weak argumentation, missing differentiation from related work, evidence-claim gaps, undefined terms.
2. Critic (writing mode) — flags overclaim, unsupported conclusions, logical gaps, hedging mismatch. Runs in parallel with the Analyst (asyncio.gather).
3. Scribe revision — produces a single revised draft that addresses both reviews while preserving your fingerprinted voice. Skips the LLM entirely if both reviews come back empty.

The console prints four panels (original draft, analyst review, critic review, revised draft); --output review.md also writes the whole bundle to a Markdown file for diffing.

research review drafts/intro.md --section introduction --interactive

Pass --interactive to walk through each reviewer issue and suggestion one at a time (y / N). The Scribe revision will only address the items you accepted, and at the end you get a coloured unified diff between original and revised. The (original, revised, selected_, rejected_) tuple is persisted to the draft_revisions SQLite table by default — pass --no-save to opt out. The continuous-learning loop (next subsection) consumes these rows to keep the fingerprint in sync with how you actually edit Scribe output.

Continuous fingerprint learning

research style update          # refresh fingerprint from samples + accepted revisions
research style history         # list all saved fingerprint versions

research style update recomputes the fingerprint by combining the static style_samples corpus with every accepted Scribe revision (revised_text from the draft_revisions table) and bumps the version. The previous fingerprint.json is archived next to it as fingerprint_v<N>.json; research style history lists every version side by side so you can see how your voice drifts as you keep using Scribe.

The flow is:

1. research style train (one-shot, from your published papers) → seeds the corpus.
2. research style fingerprint → v1 baseline.
3. research write <section> --context … → drafts.
4. research review <file> --interactive → accept / reject suggestions; the (original, revised) pair is saved.
5. research style update → folds those accepted revisions back into the fingerprint as v2, v3, …

Figure generation (M5 — research figure)

research figure --type architecture --desc "three-layer encoder with residual connections" --versions 2 --output figs/arch.md
research figure --type result       --desc "accuracy comparison across 3 baselines" --data "ours 85, baseline-A 80, baseline-B 78" --verify
research figure --type concept      --desc "attention mechanism flow" --versions 3 --output figs/concept.md

research figure drives the Illustrator agent to produce N variant drafts of a paper figure in parallel. Three modes:

• --type architecture → TikZ snippets ready to paste into LaTeX. Includes \usetikzlibrary declarations and a \tikzset{} style block. Variants cycle through layered horizontal / hub-and-spoke / encoder-decoder vertical layouts.
• --type result → matplotlib / seaborn Python scripts that write to output.png (no plt.show(), paper-ready rcParams, colorblind palette). Variants cycle through grouped bar / line with shaded variance / paired boxplot. Pass --verify to actually execute each draft in a subprocess (30 s timeout, MPLBACKEND=Agg) and report run / fail per draft.
• --type concept → text-to-image prompts for DALL·E 3, Midjourney v6, and Stable Diffusion (one variant per ecosystem, with model-specific phrasing and a negative_prompt for SD/MJ). Direct API rendering (uploading to DALL·E) is left to a follow-up.

Each draft includes a notes summary of what makes it distinct and a suggested_use phrase telling you which paper context it fits. With --output PATH the whole bouquet is written to a Markdown file with fenced code blocks and verification status; without it the output stays in the terminal as syntax-highlighted Rich panels.

Self-plagiarism check

research check drafts/intro.md --threshold 0.4 --output reports/intro.similarity.md

research check <file> scans each paragraph of the draft against every paragraph in your style_samples corpus using paragraph-level TF-IDF + cosine similarity (pure Python, no heavy dependencies). The default threshold is 0.4 (per the M4 milestone); raise it for a stricter scan or lower it to surface light echoes.

The console renders one panel summary plus a Markdown report:

• For every flagged paragraph: the draft text, the matching corpus paragraph (with arxiv:<id> source label), the similarity %, and concrete rewrite suggestions that scale with severity (≥ 0.7 → "rewrite from scratch", ≥ 0.5 → "paraphrase and cite", ≥ 0.4 → "trim or merge").
• Exit code is 0 for a clean check and 2 when at least one match trips the threshold, so it slots cleanly into CI.

Development

ruff check src tests
ruff format --check src tests
mypy src
pytest

Live network tests (hit arXiv / Semantic Scholar) are skipped by default. To run them:

RUN_NETWORK_TESTS=1 pytest -m network

Skip them explicitly with pytest -m 'not network' (already the default via RUN_NETWORK_TESTS being unset).

Planning

See planning/PLAN.md and planning/architecture.md. Milestone notes live under planning/milestones/.