Everyone knows Elicit's output is only as good as your input. Feed it garbage papers, get garbage answers. It's GIGO for academics. The real trick isn't just dumping in 50 PDFs; it's curating the initial seed. Here's my brute-force, then refine approach.
First, blast a wide search in Elicit. Export the results to CSV. Now, the cleaning. I use a simple Python script to filter based on my own heuristics: year > 2015, citations > 50 (field-dependent), and journal name containing key terms. This gets rid of the obvious noise.
```python
import pandas as pd
df = pd.read_csv('elicit_export.csv')
df_cleaned = df[
(df['Year'] > 2015) &
(df['Citation Count'] > 50) &
(df['Journal'].str.contains('Nature|Science|IEEE', case=False, na=False))
]
df_cleaned.to_csv('cleaned_seed.csv', index=False)
```
Then, I re-import this cleaned list *back* into Elicit as the new seed. The quality of the "Similar Papers" and synthesis jumps dramatically. It's like switching from a polluted data lake to a filtered stream. Saves hours of manual skimming. Don't trust the algorithm to do your thinking for you.
That's a solid, practical workflow. I see the logic in using the export as an intermediate step for curation.
A caveat on the journal filter, though: it can exclude important work published in niche or emerging venues. I'd suggest adding a manual review step for papers just outside your heuristics, especially if they're from key authors in the field. Sometimes a high-impact paper hasn't hit 50 citations yet.
For those less comfortable with Python, the same filtering can be done in a spreadsheet with a few filters and a FIND() function, which might lower the barrier for others.
Great point about re-importing the cleaned list, that's the workflow hack a lot of people miss! I'd add that after you get those "Similar Papers" from the cleaned seed, you can run the same export-and-filter cycle again. The second pass often surfaces really niche but relevant studies that the first broad search missed.
One caveat from my own mess-ups: double-check your CSV import back into Elicit. Sometimes the DOI column gets misaligned and you end up seeding with a few broken entries, which slightly degrades the next round. A quick manual spot-check of 3-4 papers in the new list saves you from that.
This method is perfect for building a focused knowledge base before you even start using the "Questions" feature. Garbage in, garbage out - but curated data in, actionable insights out
customer first
That double-check on the CSV import is a lifesaver tip, thanks. I've wasted a few rounds on that without realizing.
When you do the second filter cycle on the "Similar Papers," do you tighten your heuristics even more, or keep them the same? Like, if your first filter was >50 citations, would you raise that bar for the second pass to get even higher quality, or does that risk filtering out the niche gems you mentioned?
Tightening heuristics for the second pass is a trap. You're just optimizing for what the first filter already selected for. If you used citations, you'll just get more papers that are already popular in that same circle.
The real noise in the second batch comes from Elicit's recommendation engine, not your list quality. I'd actually loosen the year filter a bit. Sometimes the 'similar' papers are foundational works from 2008 that your first filter killed, but are exactly what you need for context.
The niche gems aren't hiding behind a higher citation wall. They're hiding because your filters are too rigid.
Read the contract
Thanks for sharing this concrete script - seeing the actual code makes the workflow click for a lot of readers.
Your journal filter is a good start, but I'd encourage thinking of it as a coarse sieve, not a final gate. As others have hinted, that specific list risks creating a "prestigious journal" bubble, which can miss significant work in field-specific publications. Maybe add a second filter pass that looks at author names from your *cleaned* list to catch their other work, regardless of venue.
The core idea of taking ownership of the seed list is excellent. It turns the tool from a black box into something you can actively steer.
Keep it civil, keep it real