Skip to content

Navigation Menu

Sign in
Sign up

Paper Wiki

Dong Dai edited this page May 26, 2026 · 2 revisions

Paper Wiki is a background agent that turns every paper you touch into a local, concept-organized knowledge base — shared across all your projects. It runs offline, is disabled by default, and is a pure enhancement: the coordinator functions identically without it.

This page is the deep-dive. For the quick version, see the README feature summary.

What it does

When you search literature, open a paper, or let the agent fetch one, Paper Wiki asynchronously:

  1. Canonicalizes identity — deduplicates across DOI, arXiv ID, and normalized title+year so the same paper isn't reprocessed under different keys.
  2. Downloads full text when available — arXiv PDFs are fetched and cached; otherwise it works from the abstract.
  3. Generates a paper page — a summary Markdown page with key claims, methodology, and extracted concepts.
  4. Synthesizes concept pages — recurring concepts get their own cross-paper synthesis page, with back-references to every paper that mentions them.
  5. Indexes for retrieval — BM25 search over paper and concept pages so the coordinator can recall earlier work via tools.

The wiki sub-agent is the single writer. The coordinator never writes to it — a cross-process lock file enforces this.

Where it lives

~/.research-pilot/paper-wiki/
├── SCHEMA.md # Wiki conventions (LLM reference)
├── index.md # Content catalog
├── log.md # Append-only operation log
├── papers/ # One .md per paper (canonical-slug keyed)
├── concepts/ # Cross-paper synthesis pages
├── raw/arxiv/ # Downloaded arXiv PDFs (cached)
├── converted/ # PDF → Markdown (cached)
└── .state/
 ├── processed.jsonl # Page-generation watermark
 ├── provenance.jsonl # Provenance tracking
 └── wiki.lock # Single-writer lock

It's global, not per-workspace — the whole point is cross-project memory. Artifacts for the current project still live under <workspace>/.research-pilot/.

Enabling it

  1. Open Settings (Cmd + .) → Paper Wiki tab.
  2. Pick a model (see cost below). Auto follows the system-wide priority (subscription first).
  3. Pick a speed preset (controls parallelism and how aggressively it indexes).

That's it. The agent starts indexing in the background the next time you interact with papers.

Cost expectations

Roughly per paper:

Tier Input Output
Abstract-only ~8K tokens ~2K tokens
Full-text (PDFs) ~25K tokens ~4K tokens

Recommendations:

  • Subscription-backed models (ChatGPT Pro / Claude Max) are cheapest — no per-token billing.
  • If using API keys: pick a fast/cheap model (e.g. Haiku, GPT-5.4 mini). The wiki doesn't need frontier reasoning.
  • Start narrow. If you're unsure, set the model to none until you've tested the app without it, then turn it on.

How the coordinator uses it

The wiki is accessed through read-only tools the agent calls automatically when relevant:

Tool Purpose
wiki_search BM25 search over paper and concept pages
wiki_get Fetch a specific page by slug
wiki_lookup Resolve a paper reference to its wiki page
wiki_coverage Report which of a set of papers are indexed
wiki_facets Query by tag / concept / venue facets
wiki_neighbors Find related papers via shared concepts
wiki_source Retrieve the raw source (PDF / abstract)

You don't call these yourself — the coordinator picks the right one based on context.

Canonical identity (why papers don't get reprocessed)

Within a project, papers are deduped by DOI > citeKey > title+year. The wiki uses a different priority for cross-project identity:

DOI > arxivId > normalized(title+year)

This means a paper you saved via arXiv in project A and via DOI in project B collapses to one wiki page. Identity drift is reconciled automatically; see lib/wiki/reconcile-identity.ts if you want the gory details.

Troubleshooting

Papers aren't being indexed

  • Confirm the wiki is enabled in Settings → Paper Wiki (model is not none).
  • Check ~/.research-pilot/paper-wiki/log.md — it's append-only and usually has the reason.
  • Make sure the chosen model has auth configured (see FAQ → Authentication).

Costs are higher than expected

  • Lower the speed preset — it controls parallelism.
  • Switch to a cheaper model. The wiki doesn't need frontier reasoning.
  • Consider abstract-only mode if most of your corpus isn't on arXiv anyway.

I want to start over

Delete ~/.research-pilot/paper-wiki/ entirely. The wiki is pure cache — your artifacts and session summaries live elsewhere and are untouched.

Design notes

Paper Wiki follows the project's design axiom:

The system does not pursue complex architecture to guarantee quality. Instead, it pursues minimum discipline to guarantee survival + evidence-driven incremental improvement.

Every failure path degrades to "wiki doesn't exist" behavior. That's why it's disabled by default and isolated from the coordinator's critical path.

For the full design, see RFC-003: Global Paper Wiki and RFC-005: Wiki Sidecar and Retrieval.

Clone this wiki locally

AltStyle によって変換されたページ (->オリジナル) /