crw-research

# fastCRW Research Find EVERY arXiv paper that answers a research query. Recall = union of arXiv ids; extra ids never hurt, so cast wide but on-topic. The [Research API](https://docs.fastcrw.com/research-api) is a live, drop-in Firecrawl-research-compatible surface — your job is **query strategy + intent routing**, the endpoints do the retrieval. This skill reaches 61.0% recall on the 191-question ArXivQA benchmark, a number fastCRW measured directly. Firecrawl publishes 53.3% for its own Research Index on the same benchmark in their [Research Index launch post](https://www.firecrawl.dev/blog/research-index-launch) (2026-06-17): "On arXivQA, the index hits 53.3% recall at $0.32 per task, against 45.4% for the next best provider." Set `FASTCRW_API_KEY` (a `crw_live_…` key from https://fastcrw.com/dashboard). Base URL `https://api.fastcrw.com`. Every endpoint is a GET; pull arXiv ids out of `results[].ids.arxiv` / `results[].primaryId`. ```bash # search: ranked papers for one query curl -s -H "Authorization: Bearer $FASTCRW_API_KEY" \ "https://api.fastcrw.com/v1/search/research/papers?query=$(jq -rn --arg q "QUERY" '$q|@uri')&k=40" # references / citers / similar of a seed paper (citation graph) curl -s -H "Authorization: Bearer $FASTCRW_API_KEY" \ "https://api.fastcrw.com/v1/search/research/papers/arxiv:1706.03762/similar?intent=related%20work&mode=references&k=40" ``` ## The whole game: classify the query, apply the matching method **A) ALWAYS (base):** write 8–12 **exact-name** queries — specific method, model, dataset, and benchmark NAMES, not broad phrases ("MoleculeNet benchmark", "Uni-Mol", "ChemBERTa", not "molecular embeddings"). Call `search` on each, union the arXiv ids, rank by how many queries surfaced each id. Exact-name decomposition is the #1 recall lever — one broad query misses the niche papers. Then **expand**: take the 5 ids the most queries agreed on and pull `/papers/arxiv:<id>/similar?mode=references` for each, appending any new ids below the union. This second pass is not optional and is not the same as (B): it runs on EVERY query type, and it is where the papers that no query names directly come from. **B) COMPARE-AGAINST** ("what does X compare to / build on / baseline against") → resolve X to its arXiv id, then `/papers/arxiv:<X>/similar?mode=references`. The answer lives in X's own bibliography. **C) USING / EXTENDING X** ("models that USE/adopt X") → `/similar?mode=citers` (forward citations) + exact-name searches for known adopters. **D) BEST-ON-BENCHMARK** ("which models score best on X", "largest open model") → search the leaderboard, read the OPEN model names (DeepSeek/Qwen/GLM/Kimi/MiniMax/Llama/Mistral/Gemma — ignore Claude/GPT/Gemini, no papers), then `search "<model family> technical report"` for each. **E) NICHE ENUMERATION** ("papers that do X") → exact-name queries (A) are primary. A tight survey or awesome-list, when on-topic, adds its ids. ## Rules - Recent ids (25xx / 26xx) are REAL — keep them, never discard as "future-dated". - A query that sounds specific usually still has a *family* of papers — surface the family, don't stop at one. Only a query naming a paper by title is single. - Merge ALL ids from every step; method-targeted (references/leaderboard) and exact-name hits first, broad-search tail after. Never invent ids.