klbr/n8nworkflows.xyz
1
0
Fork 0
mirror of https://github.com/khoaliber/n8nworkflows.xyz.git synced 2026-04-28 00:29:22 +00:00
Code Issues Packages Projects Releases Wiki Activity

creation

Browse Source
This commit is contained in:
nusquama
2025-11-13 20:10:23 +01:00
parent 3dd5cc7558
commit 693e383ba8
1 changed files with 293 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files
workflows/Automate Academic Literature Reviews with GPT-4 and Multi-Database Search-8503/readme-8503.md
+293
View File
@@ -0,0 +1,293 @@
Automate Academic Literature Reviews with GPT-4 and Multi-Database Search
https://n8nworkflows.xyz/workflows/automate-academic-literature-reviews-with-gpt-4-and-multi-database-search-8503
# Automate Academic Literature Reviews with GPT-4 and Multi-Database Search
### 1. Workflow Overview
This workflow automates academic literature reviews using GPT-4 and multi-database search capabilities. It targets researchers and systematic reviewers wishing to streamline the process of collecting, screening, analyzing, and synthesizing academic papers on a specific topic. The workflow logically divides into the following blocks:
- **1.1 Review Setup:** Defines search parameters for the literature review, such as topic and publication year range.
- **1.2 Multi-Database Search:** Searches multiple academic databases (PubMed, Semantic Scholar, arXiv) for relevant papers based on the defined parameters and retrieves metadata including PDFs when available.
- **1.3 Paper Ranking & Selection:** Ranks retrieved papers by relevance and citation metrics, selecting the top N papers for further processing.
- **1.4 Paper Processing Loop:** Processes each selected paper individually, checking for PDFs, parsing content if available, or falling back to abstracts.
- **1.5 AI Analysis:** Uses GPT-4 to generate detailed review entries for each paper including summaries, methodologies, findings, and citations.
- **1.6 Review Compilation:** Aggregates all review entries, categorizes them into thematic groups, and compiles a final literature review document with summaries and a bibliography.
Sticky notes provide contextual guidance and explain key parts of the workflow: overview, search strategy, quality assessment criteria, and final outputs.
---
### 2. Block-by-Block Analysis
#### 2.1 Review Setup
- **Overview:** Defines the literature review’s scope by setting parameters like topic keywords, year range, and maximum number of papers to process.
- **Nodes Involved:**
- Set Search Parameters
- Review Overview (Sticky Note)
- **Node Details:**
- **Set Search Parameters**
- Type: Set Node
- Role: Configures input variables for the workflow: topic ("machine learning in healthcare"), year range (2020–2024), and max papers (20).
- Expressions: Directly sets static values, no dynamic expressions.
- Inputs: None (start node)
- Outputs: Connects to PDF Vector - Search Papers
- Edge Cases: Incorrect parameter values may limit or skew search results; no validation logic included.
- Notes: Sticky note “Review Overview” contextualizes this block’s role.
- **Review Overview (Sticky Note)**
- Type: Sticky Note
- Provides high-level description of the review automation goals: search, screen, quality assess, synthesize, generate.
#### 2.2 Multi-Database Search
- **Overview:** Executes a multi-provider search on academic databases using the query parameters; retrieves paper metadata including abstracts and PDFs.
- **Nodes Involved:**
- PDF Vector - Search Papers
- Database Search (Sticky Note)
- **Node Details:**
- **PDF Vector - Search Papers**
- Type: PDF Vector Node (custom academic search node)
- Role: Queries PubMed, Semantic Scholar, and arXiv with the topic and year filter; retrieves up to 50 papers with fields like title, abstract, authors, year, DOI, PDF URL, citations.
- Expressions: Parameters dynamically use values from ‘Set Search Parameters’ node for topic, yearFrom, yearTo.
- Inputs: From Set Search Parameters
- Outputs: To Rank & Select Papers
- Edge Cases: API limits, missing metadata, or empty results; failure to retrieve PDFs could limit later steps.
- Notes: Sticky note “Database Search” highlights covered databases and de-duplication.
#### 2.3 Paper Ranking & Selection
- **Overview:** Scores and ranks retrieved papers by combining citation counts, recency, and text relevance to the topic; limits to a maximum number.
- **Nodes Involved:**
- Rank & Select Papers
- Process One by One (split batch)
- **Node Details:**
- **Rank & Select Papers**
- Type: Code Node (JavaScript)
- Role: Implements custom logic to compute relevance scores: citation score (40%), recency score (up to 20 points), title and abstract keyword matches (30 points total). Sorts papers by score and selects top N (maxPapers).
- Expressions: Reads parameters from ‘Set Search Parameters’ and paper metadata from input.
- Inputs: From PDF Vector - Search Papers
- Outputs: To Process One by One
- Edge Cases: Zero citations handling, incomplete metadata, or no keyword matches could affect ranking fairness. Code assumes numeric year parsing.
- Notes: Processes all retrieved papers collectively before batching.
- **Process One by One**
- Type: SplitInBatches Node
- Role: Converts the ranked list into individual paper items for sequential processing downstream.
- Inputs: From Rank & Select Papers
- Outputs: To Has PDF?
- Edge Cases: Batch size fixed to 1; large lists could slow workflow.
#### 2.4 Paper Processing Loop
- **Overview:** For each paper, checks availability of PDF; if available, parses it to extract content; otherwise, proceeds with abstract.
- **Nodes Involved:**
- Has PDF?
- PDF Vector - Parse Paper
- Analyze Paper Content
- Store Review Entry
- **Node Details:**
- **Has PDF?**
- Type: If Node
- Role: Conditional branch based on whether paper metadata contains a non-empty PDF URL.
- Expressions: Checks if `pdfURL` field is not empty.
- Inputs: From Process One by One
- Outputs:
- True: PDF Vector - Parse Paper
- False: Analyze Paper Content (directly uses abstract content)
- Edge Cases: Missing or broken URLs may cause false negatives; no fallback for inaccessible PDFs.
- **PDF Vector - Parse Paper**
- Type: PDF Vector Node
- Role: Parses PDF content from URL; supports auto LLM usage for content extraction.
- Inputs: From Has PDF? (true branch)
- Outputs: To Analyze Paper Content
- Edge Cases: PDF parsing failures, timeouts, or unreadable PDFs may cause errors or incomplete content.
- **Analyze Paper Content**
- Type: OpenAI Node (GPT-4)
- Role: Sends paper metadata and content (PDF-extracted or abstract) to GPT-4 to generate a structured literature review entry with summary, methodology, findings, topic relevance, limitations, and APA citation.
- Configuration: Uses a detailed prompt template referencing ‘Set Search Parameters.topic’ and paper fields.
- Inputs: From PDF Vector - Parse Paper or Has PDF? (false branch)
- Outputs: To Store Review Entry
- Edge Cases: API quota limits, rate limiting, or prompt errors may cause failures or incomplete responses.
- **Store Review Entry**
- Type: Set Node
- Role: Extracts AI-generated review text and stores it along with paper title and DOI in variables for later compilation.
- Expressions: Reads GPT-4 output JSON to extract review content and paper metadata from condition node.
- Inputs: From Analyze Paper Content
- Outputs: To Process One by One (to continue batch processing)
- Edge Cases: Missing fields in GPT response could lead to empty entries.
#### 2.5 Review Compilation
- **Overview:** After processing all papers, aggregates and categorizes review entries into thematic groups; generates a comprehensive literature review document with summaries and bibliographic references.
- **Nodes Involved:**
- Compile Literature Review
- Final Review (Sticky Note)
- **Node Details:**
- **Compile Literature Review**
- Type: Code Node (JavaScript)
- Role: Collects all stored review entries, groups them into predefined themes (e.g., Machine Learning Models, Clinical Applications) based on keyword matching in review text, and constructs a Markdown-formatted literature review document. Also compiles a bibliography from suggested citations.
- Inputs: Aggregates all stored review entries after batch completion.
- Outputs: Final literature review document with metadata (total papers, theme counts, generation timestamp).
- Edge Cases: Thematic grouping is heuristic and simplistic; complex NLP categorization is recommended for production. Missing citations handled gracefully with placeholder text.
- Notes: Sticky note “Final Review” describes output contents and formatting.
---
### 3. Summary Table
| Node Name | Node Type | Functional Role | Input Node(s) | Output Node(s) | Sticky Note |
|-------------------------|--------------------------|---------------------------------|------------------------|---------------------------|--------------------------------------------------------------------------------------------------------------|
| Review Overview | Sticky Note | Workflow purpose and scope | — | — | ## 📖 Literature Review Generator Systematic review automation: searches, screens, assesses, synthesizes, generates PRISMA-compliant review |
| Set Search Parameters | Set | Define review search parameters | — | PDF Vector - Search Papers | Configure literature review parameters |
| PDF Vector - Search Papers | PDF Vector Node | Search academic databases | Set Search Parameters | Rank & Select Papers | ## 🔍 Search Strategy Databases searched: PubMed/MEDLINE, Web of Science, Cochrane Library, Google Scholar De-duplicates results |
| Rank & Select Papers | Code | Rank papers by relevance | PDF Vector - Search Papers | Process One by One | Rank papers by relevance |
| Process One by One | SplitInBatches | Batch processing of papers | Rank & Select Papers | Has PDF? | Process papers individually |
| Has PDF? | If | Check for PDF availability | Process One by One | PDF Vector - Parse Paper (true), Analyze Paper Content (false) | |
| PDF Vector - Parse Paper | PDF Vector Node | Parse paper content from PDF | Has PDF? (true) | Analyze Paper Content | Parse paper content from PDF or image |
| Analyze Paper Content | OpenAI (GPT-4) | Generate review entry via AI | PDF Vector - Parse Paper, Has PDF? (false) | Store Review Entry | Generate review entry |
| Store Review Entry | Set | Store AI-generated review entry | Analyze Paper Content | Process One by One | Save processed entry |
| Compile Literature Review | Code | Compile final literature review | Store Review Entry (aggregated) | — | Generate final document |
| Final Review | Sticky Note | Describes final review outputs | — | — | ## 📝 Review Output Generates narrative synthesis, evidence tables, PRISMA diagram, forest plots, bibliography Publication ready! |
---
### 4. Reproducing the Workflow from Scratch
1. **Create Sticky Notes for Context:**
- Add a sticky note titled "Review Overview" at the start with content describing the workflow’s purpose: automated systematic review generation.
- Add sticky notes "Database Search," "Study Quality," and "Final Review" near respective logical blocks for user guidance.
2. **Set Search Parameters Node:**
- Add a **Set** node named "Set Search Parameters."
- Configure variables:
- `maxPapers` (number): 20
- `topic` (string): "machine learning in healthcare"
- `yearFrom` (string): "2020"
- `yearTo` (string): "2024"
- No inputs; connect output to next node.
3. **PDF Vector - Search Papers Node:**
- Add **PDF Vector** node named "PDF Vector - Search Papers."
- Configure:
- Operation: `search`
- Resource: `academic`
- Query: `={{ $json.topic }}`
- Year From: `={{ $json.yearFrom }}`
- Year To: `={{ $json.yearTo }}`
- Providers: select `pubmed`, `semantic-scholar`, `arxiv`
- Limit: 50
- Additional fields: title, abstract, authors, year, doi, pdfURL, totalCitations
- Connect input from "Set Search Parameters."
4. **Rank & Select Papers Node:**
- Add a **Code** node named "Rank & Select Papers."
- Paste the JavaScript code that:
- Collects papers from input
- Calculates relevance scores based on citations, recency, title and abstract keyword matches
- Sorts and selects top N (`maxPapers`) papers
- Connect input from "PDF Vector - Search Papers."
5. **Process One by One Node:**
- Add **SplitInBatches** node named "Process One by One."
- Set batch size to 1.
- Connect input from "Rank & Select Papers."
6. **Has PDF? Node:**
- Add **If** node named "Has PDF?"
- Condition: Check if `pdfURL` field is not empty (`isNotEmpty`).
- Connect input from "Process One by One."
- True output connects to "PDF Vector - Parse Paper."
- False output connects to "Analyze Paper Content."
7. **PDF Vector - Parse Paper Node:**
- Add **PDF Vector** node named "PDF Vector - Parse Paper."
- Configure:
- Operation: `parse`
- Resource: `document`
- Input Type: `url`
- URL: `={{ $json.pdfURL }}`
- Use LLM: `auto`
- Connect input from "Has PDF?" true branch.
- Output connects to "Analyze Paper Content."
8. **Analyze Paper Content Node:**
- Add **OpenAI** node named "Analyze Paper Content."
- Set model to `gpt-4`.
- Configure prompt message to:
```
Create a literature review entry for this paper in the context of '{{ $node['Set Search Parameters'].json.topic }}':
Title: {{ $json.title }}
Authors: {{ $json.authors }}
Year: {{ $json.year }}
Citations: {{ $json.totalCitations }}
Content: {{ $json.content || $json.abstract }}
Provide:
1. A 3-4 sentence summary of the paper's contribution
2. Key methodology used
3. Main findings (2-3 bullet points)
4. How it relates to the topic
5. Limitations mentioned
6. Suggested citation in APA format
```
- Connect input from "PDF Vector - Parse Paper" and "Has PDF?" false branch.
9. **Store Review Entry Node:**
- Add **Set** node named "Store Review Entry."
- Configure to save:
- `reviewEntry` from GPT-4 output: `={{ $json.choices[0].message.content }}`
- `paperTitle` from condition node: `={{ $node['Has PDF?'].json.title }}`
- `paperDoi` from condition node: `={{ $node['Has PDF?'].json.doi }}`
- Connect input from "Analyze Paper Content."
- Output connects back to "Process One by One" to continue batching.
10. **Compile Literature Review Node:**
- Add **Code** node named "Compile Literature Review."
- Paste JavaScript that:
- Collects all review entries
- Categorizes them into themes based on keywords in review text
- Builds a Markdown document with sections per theme, summaries, and full bibliography
- Connect input from "Store Review Entry" (all completed batches).
11. **Final Review Sticky Note:**
- Add sticky note near compilation node describing final outputs: narrative synthesis, evidence tables, PRISMA diagram, forest plots, bibliography, publication-ready document.
12. **Credentials Setup:**
- Configure OpenAI credentials with model access to GPT-4.
- Ensure PDF Vector node credentials or API keys for academic database access are configured.
- No explicit OAuth nodes shown; ensure access tokens or API keys are valid.
---
### 5. General Notes & Resources
| Note Content | Context or Link |
|----------------------------------------------------------------------------------------------|-------------------------------------------------------|
| Workflow automates systematic literature reviews with multi-database search and GPT-4 analysis. | Overall workflow purpose |
| Uses PDF Vector node for both academic search and PDF content parsing, leveraging LLM where needed. | PDF Vector node documentation: https://docs.pdfvector.com/ (hypothetical) |
| Ranking algorithm balances citations, recency, and keyword relevance for paper selection. | Custom ranking logic embedded in code node |
| Thematic grouping in final compilation is heuristic; consider integrating advanced NLP for production. | Potential enhancement area |
| Final output includes PRISMA-compliant elements, evidence tables, and bibliographic formatting. | Supports publication-ready systematic reviews |
---
This documentation fully describes the workflow's architecture, node configurations, data flows, and logic, enabling accurate reproduction or modification by users or AI agents.