Paperscope

Search, explore, and chat with ~6,000 AI research papers from arXiv, powered by Meilisearch.

Live demo: paperscope-one.vercel.app

Type a title, an author, or describe an idea in your own words. A slider moves the search from exact words to meaning, every paper links to its nearest neighbours, and the chat page answers questions from the whole corpus, your reading list, or a single paper, with sources.

What it shows

In the app	Meilisearch feature
"Exact words ↔ Meaning" slider	Hybrid search (hybrid.semanticRatio) with a HuggingFace embedder (BAAI/bge-small-en-v1.5) running inside Meilisearch
"11 of them found only by meaning"	semanticHitCount
"100% match" on each result	showRankingScore
"tranformer" still finds transformers	Typo tolerance
"LLM", "RAG", "LoRA"...	Synonyms
Category, topic, and author filters with counts	Filters and facets
Year range slider and "Results span 2003 to 2026"	facetStats
"Find an author..." box	Facet search
Sort by relevance, most cited, newest, oldest	Sort, plus the custom ranking rule citationCount:desc
Highlighted matches and abstract snippets	Highlighting and cropping
Author chips above the results	Multi-search (papers and authors indexes in one request)
"Similar papers" in the paper panel	/similar endpoint, optionally filtered to the same category
"Ask the papers" page	Conversational search (/chats): the LLM calls a hybrid search tool and streams progress and sources
Chat with your reading list or one paper	Tenant token whose search rule is id IN [...]

Run it locally

Requires Docker (OrbStack) and Node 24.

cp .env.example .env               # add CHAT_API_KEY to enable the chat page
docker compose up -d meilisearch
cd web && pnpm install
pnpm data:fetch                    # arXiv + Semantic Scholar -> data/papers.json (~5 min)
pnpm data:setup                    # settings, embeddings, chat workspace, API keys (~15 min on CPU)
cd .. && docker compose watch      # web app with live sync

• App: http://web.research-papers.orb.local:3000 (or http://localhost:3100)
• Meilisearch: http://meilisearch.research-papers.orb.local:7700 (or http://localhost:7700), master key in .env

pnpm data:setup prints the search-only key. The one in .env.example already matches the default master key.

Chat LLM

The /chats workspace needs an LLM provider. Set these in .env, then re-run pnpm data:setup (documents are already indexed, so it is fast):

CHAT_SOURCE=openAi        # openAi | mistral | vLlm
CHAT_API_KEY=sk-...
CHAT_MODEL=gpt-4o-mini
CHAT_BASE_URL=            # required for mistral (https://api.mistral.ai/v1) and vLlm

Without it, search and similar papers still work; only the chat page is disabled.

Dataset size

PER_TOPIC=50 RECENT_PER_CATEGORY=100 pnpm data:fetch gives ~2,000 papers. Topics and categories are listed in web/scripts/fetch-arxiv.ts.

How it works

The browser queries Meilisearch directly with a search-only key: multi-search, /similar, and facet search all live in web/src/lib/search-api.ts. The only server route that matters is /api/chat-token, which signs a 30-minute tenant token so the browser can call /chats without ever seeing the chat key. The token's search rules restrict the LLM to the papers in scope.

All Meilisearch configuration (indexes, settings, synonyms, embedder, chat workspace, keys) lives in one script, web/scripts/setup-meilisearch.ts, and is safe to re-run.

compose.yaml                      meilisearch + web (Next.js dev server, compose watch)
data/papers.json                  fetched dataset (git-ignored)
web/scripts/fetch-arxiv.ts        dataset builder (arXiv API, Semantic Scholar citations and TL;DRs)
web/scripts/setup-meilisearch.ts  all Meilisearch configuration
web/src/lib/search-api.ts         browser-side search (search-only key)
web/src/app/api/chat-token        signs tenant tokens for /chats
web/src/app/api/status            health check
web/src/app/                      UI: search page, paper panel, chat page

Production

Piece	Where
Front	Vercel, team meili, project paperscope: https://paperscope-one.vercel.app
Meilisearch	Main instance on qdq-server (v1.54), papers and authors indexes, at https://search.qdq.meilisearch.com
Chat LLM	LUMEN on the same box, through its public URL https://lumen.meilisearch.com/v1, model claude-sonnet-4-5, virtual key paperscope-demo-chat ($25 hard budget, 60 rpm)

Two things worth knowing:

• Search is not proxied through Vercel. Caddy rate-limits per client IP, so a proxy would put every visitor in the same bucket.
• Meilisearch rejects private IPs such as 127.0.0.1 as a chat baseUrl, which is why LUMEN is reached through its public URL even though it runs on the same box.

Keys are created by setup-meilisearch.ts with fixed uids, so re-runs find them again:

Key	uid	Actions	Indexes	Lives in
Paperscope search (public)	7dbdf598-3e28-4236-9e10-3a5032b8ee19	search	papers, authors	NEXT_PUBLIC_MEILI_SEARCH_KEY (browser bundle)
Paperscope chat (server)	aaea7803-0890-4396-b0e5-1a630e1a3dc9	search, chatCompletions	papers	MEILI_CHAT_KEY and MEILI_CHAT_KEY_UID (Vercel, server only)

The LUMEN key is stored on the box in /etc/meilisearch/paperscope-lumen-key.json (0600) and in the papers chat workspace settings.

Re-import

Run the setup over an SSH tunnel, copying embeddings from your local instance so the box does not compute them:

ssh -f -N -L 17700:127.0.0.1:7700 root@62.210.158.50
cd web
MEILI_HOST=http://localhost:17700 \
MEILI_MASTER_KEY="$(ssh root@62.210.158.50 'sed -n "s/^MEILI_MASTER_KEY=//p" /etc/meilisearch/meilisearch.env')" \
CHAT_API_KEY="$(ssh root@62.210.158.50 'python3 -c "import json;print(json.load(open(\"/etc/meilisearch/paperscope-lumen-key.json\"))[\"key\"])"')" \
CHAT_BASE_URL=https://lumen.meilisearch.com/v1 \
VECTORS_FROM_HOST=http://localhost:7700 \
node scripts/setup-meilisearch.ts

Deploy the front

cd web && vercel deploy --prod --scope meili