Knowledge Pipeline Skill

Metadata

• Name: knowledge-pipeline
• Version: 1.1
• Purpose: Build rich knowledge context using RAG + PubMed before writing
• Trigger: Any content creation task requiring evidence-based writing

Overview

This skill implements a parallel knowledge building pipeline that queries BOTH:

1. RAG Pipeline - Your AstraDB vector store containing cardiology textbooks and guidelines
2. PubMed Pipeline - Latest research via NCBI E-utilities API

NOTE: Perplexity is used SEPARATELY for social listening and demand assessment (YouTube workflow), NOT for research/evidence gathering.

When to Use

ALWAYS use this skill BEFORE writing content that requires:

• Evidence-based claims
• Statistics or study citations
• Guideline references
• Current best practices
• Recent trial results

Architecture

┌─────────────────────────────────────────────────────────────┐
│                 KNOWLEDGE PIPELINE (Research)                │
├─────────────────────────────────────────────────────────────┤
│                                                              │
│  Question ──┬──► RAG Pipeline ──────────────────────────────│
│             │    (AstraDB Vector Store)                      │
│             │    - YOUR textbooks (Braunwald, etc.)          │
│             │    - Guidelines (ESC, ACC, AHA)                │
│             │    - Reference materials                       │
│             │    Tech: Vector + BM25 + RRF + Cohere rerank   │
│             │                                                │
│             └──► PubMed Pipeline ───────────────────────────│
│                  (NCBI E-utilities API)                      │
│                  - Latest research articles                  │
│                  - Systematic reviews                        │
│                  - Meta-analyses                             │
│                  - Clinical trials                           │
│                                                              │
├─────────────────────────────────────────────────────────────┤
│                       SYNTHESIS                              │
│  Combined context ──► GPT-4o-mini ──► Knowledge Brief       │
│                                                              │
│  Output:                                                     │
│  1. Established Knowledge (guidelines)                       │
│  2. Latest Research (PubMed)                                 │
│  3. Key Data Points                                          │
│  4. Areas of Consensus                                       │
│  5. Areas of Uncertainty                                     │
│  6. Citation Summary                                         │
└─────────────────────────────────────────────────────────────┘

┌─────────────────────────────────────────────────────────────┐
│              SEPARATE: DEMAND ASSESSMENT (YouTube)           │
├─────────────────────────────────────────────────────────────┤
│  Perplexity ──► Social listening, trends, what people ask   │
│  Free LLMs  ──► Demand analysis from YouTube comments       │
│                                                              │
│  This is NOT research. This is audience intelligence.        │
└─────────────────────────────────────────────────────────────┘

Usage

Python Integration

from rag_pipeline.src.knowledge_pipeline import KnowledgePipeline

pipeline = KnowledgePipeline(verbose=True)

# Option 1: Get raw combined context
context = pipeline.build_knowledge_context(
    "What are optimal LDL targets for high-risk patients?"
)

# Option 2: Get synthesized knowledge brief
brief = pipeline.synthesize_knowledge(
    "What are optimal LDL targets for high-risk patients?"
)

CLI Usage

cd "/Users/shaileshsingh/cowriting system/rag-pipeline"
python src/knowledge_pipeline.py

Configuration

Environment Variables (in .env)

# RAG Pipeline (AstraDB)
ASTRA_DB_APPLICATION_TOKEN=your_token
ASTRA_DB_API_ENDPOINT=your_endpoint
ASTRA_DB_COLLECTION=documents
OPENAI_API_KEY=your_key
COHERE_API_KEY=your_key

# PubMed Pipeline
NCBI_API_key=your_key

Output Format

Raw Context (build_knowledge_context)

## FROM TEXTBOOKS & GUIDELINES (RAG)
--------------------------------------------------
[Source 1: ESC Guidelines 2021.pdf, Page 45] (Score: 0.892)
LDL-C targets for patients with established CVD...

[Source 2: Braunwald Cardiology.pdf, Page 1203] (Score: 0.856)
The evidence for aggressive LDL lowering...

## FROM PUBMED (Latest Research)
--------------------------------------------------
[1] PMID: 38123456
    Smith, Jones, Brown et al. (2024)
    Novel LDL-C targets in high-risk populations
    Journal of the American College of Cardiology
    Abstract: Recent meta-analysis of 15 trials...

==================================================
KNOWLEDGE SUMMARY
- RAG chunks (textbooks/guidelines): 8
- PubMed articles (latest research): 5
- Total sources: 13
==================================================

Priority When Sources Conflict

1. Highest: Established guidelines from RAG (ESC, ACC, AHA)
2. High: Major trials and meta-analyses (RAG + PubMed)
3. Medium: Recent updates not yet in guidelines (PubMed)
4. Lower: Single studies, expert opinion

Integration with Writing Skills

This skill feeds into ALL writing skills:

1. cardiology-writer - Uses knowledge brief for factual grounding
2. cardiology-newsletter-writer - Research phase uses this pipeline
3. cardiology-editorial - Evidence synthesis from both sources
4. youtube-script-master - Educational sections use RAG + PubMed context

Cost Estimates

Component	Model/Service	Cost per Query
RAG Embeddings	text-embedding-3-small	~$0.001
RAG Reranking	Cohere rerank-english-v3.0	~$0.01
PubMed API	NCBI E-utilities	Free
Synthesis	GPT-4o-mini	~$0.002
Total		~$0.013/query

Maintenance

Adding New Documents to RAG

cd "/Users/shaileshsingh/cowriting system/rag-pipeline"
python src/ingest_documents.py --folder /path/to/new/pdfs

Rebuilding BM25 Index

Delete cache files to force rebuild:

rm .bm25_cache.pkl .doc_cache.json