Karte 09 · Kapitel tools
Your own literature review pipeline
🟢 USE — Run first
0 - 15 min
Run a literature review on a topic you actually care about
You have already seen this pipeline in Session 8 — esearch → efetch → AI extract → filter → format. Today you run it for real, against a question from your own field, and judge whether the output is good enough to trust.
- Go to Downloads (curriculum.32dots.de/share) and re-import 'Session 8 — Multi-stage literature pipeline' if it's not already in your n8n.
- Pick a narrow scientific question from your own work — e.g. 'CRISPR screens in primary T cells', 'gut microbiome and Parkinson's', 'single-cell RNA-seq of glioblastoma'. Specific beats broad.
- Open the first node and change the PubMed query to your question. Keep retmax at 8 for the first run.
- Execute the workflow. Watch the structured extraction column-by-column: method, sample size, key finding, limitation.
- Read three of the rows critically. Find at least one place where the AI got the method wrong, missed a limitation, or invented a number.
Done-Signal: You can name one paper where the AI extraction is good, one where it's wrong, and one where it's plausibly wrong but you'd need to read the full paper to tell.
🔵 UNDERSTAND — Look inside
15 - 60 min
Why this pipeline beats a single prompt — and where it still fails
The five-stage shape is not decorative. Each stage exists because doing the whole thing in one LLM call produces hallucinated PMIDs, dropped limitations, and inconsistent JSON. Look at each stage and ask: what would break if I removed it?
🔎 esearch (PubMed)
HTTP — search for IDs
Returns PMIDs only. Cheap, deterministic, and grounds every downstream claim in a real article ID. No LLM here on purpose.
📄 efetch (PubMed)
HTTP — fetch abstracts
Pulls the actual abstract text. The LLM never invents PMIDs because it never sees a list to invent from — it only summarises text it was handed.
🧠 AI Extract (structured)
AI Agent — JSON output
Extracts method / sample size / key finding / limitation as structured JSON. JSON schema constrains the shape so downstream code can rely on the keys.
🧹 Filter (Code)
Code — drop incomplete rows
Throws out rows missing method or sample size. Deterministic, fast, no token cost. The earlier you filter, the cheaper the pipeline.
📊 Format (Code)
Code — render table
Renders a Markdown comparison table. Pure deterministic transform of structured data — never goes back to an LLM.
In a science pipeline, the AI is one stage of five, not the whole thing. Search and fetch are deterministic so PMIDs are real. Filter and format are deterministic so the output is reproducible. The LLM only does the one thing a deterministic step cannot — read prose and pull out structure. This is the pattern you will see in every serious research-AI system.
Probe-Fragen
- If you remove the Filter node, what kind of garbage shows up in the table?
- Why is the extraction step the only AI step? What goes wrong if you let the AI also choose which papers to include?
- Where would you add a second LLM call — and what would justify the extra cost?
🟠 BUILD — Make it yours
60 - 90 min
Tighten the pipeline for your domain
A generic pipeline is a demo. A pipeline tuned to your field is a tool. Pick one of these extensions and ship it.
Aufgabe: Choose ONE concrete extension and implement it end-to-end. Keep the change small — one node added or modified — and test that it works on at least three papers.
- Option A — Add a domain filter. In the Filter node, additionally drop rows where method does not contain one of your domain's keywords (e.g. 'qPCR', 'flow cytometry', 'scRNA-seq', 'mass spec'). Aim for precision over recall.
- Option B — Add a 'sample size' threshold. Drop rows where sample size is below a number that makes sense for your field (n<10 for cell-line work, n<50 for clinical observational, etc.). Decide the cutoff before you look at the results.
- Option C — Add a second extraction field. Pick one: 'cell line / organism', 'statistical test', 'effect size', 'control group description'. Update the JSON schema in the AI Extract node and the table header in the Format node.
- Option D — Replace the AI Extract prompt with a stricter version that refuses to answer when the abstract does not contain the requested field — and verify it actually refuses on a few abstracts where the field is genuinely absent.
- Run the modified pipeline on the same query you used in the USE phase. Compare the before/after tables side by side.
Deliverable: Export the modified workflow JSON. Share the before/after comparison table and one sentence on what you would still not trust this pipeline to do for your real research.
✓ SELF-CHECK
Hast du das verstanden?
- I ran the pipeline against a question from my own field, not the demo one.
- I found at least one extraction error and can describe what went wrong.
- I can explain why search and filter are deterministic and extraction is the only AI step.
- I implemented one extension and verified it changes the output on three papers.
This pipeline reads abstracts, not papers. For your field, when is an abstract enough — and when does relying on abstracts actively mislead a literature review?
💬 KI-TUTOR
Frag den Tutor zu dieser Karte
Sokratisch: der Tutor antwortet mit Leitfragen statt fertigen Antworten — du erarbeitest die Lösung selbst.
Stell eine erste Frage zu dieser Karte unten.