Falcon Deep Literature

Falcon — Deep Scientific Literature Synthesis (FutureHouse Platform)

Falcon is FutureHouse's high-reasoning literature agent. Same retrieval backbone as Crow (PaperQA2 on Semantic Scholar / Crossref / Unpaywall), but using LITERATURE_HIGH mode which enables extended reasoning, broader paper coverage, and produces synthesis-quality answers suitable for review-style writing.

Use Falcon when the user asks for:

• A thematic review of a topic ("Summarize the state of evidence on …")
• Gap analysis ("What's missing from the literature on …")
• Comparison across mechanisms / drugs / methods
• Background section drafting for a paper

For one-shot factual questions, Crow is faster and cheaper — only escalate to Falcon when depth matters.

Prerequisites

• pip install edison-client
• EDISON_API_KEY from https://platform.edisonscientific.com/profile

Minimal usage

import os
from edison_client import EdisonClient, JobNames

client = EdisonClient(api_key=os.environ["EDISON_API_KEY"])
resp = client.run_tasks_until_done({
    "name": JobNames.LITERATURE_HIGH,
    "query": (
        "Synthesize the current evidence for and against using GLP-1 receptor "
        "agonists in the treatment of non-alcoholic steatohepatitis (NASH). "
        "Cover mechanism, key clinical trials, head-to-head comparisons with "
        "other therapies, and remaining uncertainties."
    ),
})

print(resp.formatted_answer)

formatted_answer contains the synthesis with inline citations and a sorted reference list. Use it directly as the body of a "Background" or "Related Work" section.

Recipes

Multi-section review by parallel queries

Falcon answers one question at a time. For a multi-section review, fire several queries in parallel and stitch:

import asyncio
from edison_client import EdisonClient, JobNames

sections = {
    "Mechanism":   "Summarize the molecular mechanism of GLP-1RA action in hepatocytes.",
    "Trials":      "Review phase 2 and 3 RCTs of GLP-1RAs in NASH.",
    "Comparators": "Compare GLP-1RA efficacy in NASH vs FGF21 analogs and pioglitazone.",
    "Gaps":        "What aspects of GLP-1RA use in NASH remain underexplored?",
}

async def main():
    client = EdisonClient(api_key=os.environ["EDISON_API_KEY"])
    tasks = [{"name": JobNames.LITERATURE_HIGH, "query": q} for q in sections.values()]
    answers = await client.arun_tasks_until_done(tasks)
    for h, a in zip(sections, answers):
        print(f"## {h}\n\n{a.formatted_answer}\n")

asyncio.run(main())

Increase max steps / timeout for very deep synthesis

from edison_client.models.app import TaskRequest

resp = client.run_tasks_until_done(TaskRequest(
    name=JobNames.LITERATURE_HIGH,
    query="...",
    runtime_config={"max_steps": 60, "timeout": 1800},
))

Verbose mode — get retrieved papers + intermediate state

resp = client.run_tasks_until_done(
    {"name": JobNames.LITERATURE_HIGH, "query": "..."},
    verbose=True,
)
# resp.environment_frame contains contexts, paper metadata, scored chunks
papers = resp.environment_frame["docs"]

Cost / latency expectations

• Latency: 1–8 minutes per query (vs Crow's 20–90 s)
• Credits: roughly 5–10× a Crow call
• Quality: SOTA on LitQA2 / WikiCrow benchmarks; outputs are typically usable in published reviews after light editing

When to fall back to Crow

If your question is fact-shaped ("what dose", "what year", "did paper X show Y"), call Crow instead — Falcon's extra reasoning budget is wasted and you'll burn 10× the credits for the same answer.