Update README.md

README.md

@@ -16,26 +16,48 @@ tags:
 - chromadb
 - vector-database
 - sentence-transformers
+- bge
 size_categories:
-- 1M<n<10M
-pretty_name: STXBP1 RAG Database - ARIA
+- 100K<n<1M
+pretty_name: STXBP1 RAG Database v9 - BGE Embeddings
 ---
 
-# 🧬 STXBP1-ARIA RAG Database - (597 Million Tokens)
+# 🧬 STXBP1-ARIA RAG Database v9 - BGE Embeddings
 
-A pre-built ChromaDB vector database containing **1,194,693 indexed text chunks** from **31,786 PubMed Central (PMC) biomedical papers** (5947 Million indexed Tokens <100ms) related to STXBP1, Munc18-1, synaptic transmission, epileptic encephalopathy, and related therapeutic research.
+A pre-built ChromaDB vector database containing **~570,000 indexed text chunks** from **~17,000 curated PubMed Central (PMC) biomedical papers** related to STXBP1, Munc18-1, synaptic transmission, epileptic encephalopathy, and therapeutic research.
+
+> 💡 **This is the lightweight version** — BGE-base runs efficiently on **CPU/system RAM** without requiring a GPU, making it ideal for free-tier deployments and local development. For maximum retrieval quality with NVIDIA's state-of-the-art 2048-dimensional embeddings (requires GPU with 2-4GB VRAM), see our premium database: **[STXBP1-RAG-Nemotron](https://huggingface.co/datasets/SkyWhal3/STXBP1-RAG-Nemotron)**
+
+## 🆕 What's New in v9
+
+| Feature | v8 (Previous) | v9 (Current) |
+|---------|---------------|--------------|
+| **Embedding Model** | all-MiniLM-L6-v2 | **BGE-base-en-v1.5** |
+| **Dimensions** | 384 | **768** |
+| **Model Params** | 22M | **110M** |
+| **MTEB Score** | ~56 | **~63** |
+| **Corpus** | 31,786 papers (unfiltered) | **~17,000 papers (curated)** |
+| **Chunks** | 1.19M (58% noise) | **~570K (high relevance)** |
+| **Quality Focus** | Quantity | **Precision** |
+
+### Why BGE?
+
+- **2x embedding dimensions** = finer semantic distinctions
+- **5x larger model** = better understanding of biomedical terminology  
+- **Curated corpus** = removed irrelevant papers, kept STXBP1-focused content
+- **MTEB benchmark leader** = proven retrieval performance
 
 ## 📊 Dataset Statistics
 
 | Metric | Value |
 |--------|-------|
-| **Total Chunks** | 1,194,693 |
-| **Source Papers** | 31,786 PMC articles |
-| **Database Size** | ~17 GB |
-| **Embedding Model** | `all-MiniLM-L6-v2` (384 dimensions) |
-| **Chunk Size** | ~500 tokens with overlap |
-| **Token Count** | ~597,346,500 tokens |
+| **Total Chunks** | ~570,000 |
+| **Source Papers** | ~17,000 PMC articles |
+| **Database Size** | ~8-10 GB |
+| **Embedding Model** | `BAAI/bge-base-en-v1.5` (768 dimensions) |
+| **Chunk Size** | ~1500 chars with 200 char overlap |
 | **Index Type** | ChromaDB with HNSW |
+| **Build Date** | January 2026 |
 
 ## 🎯 Purpose
 
@@ -50,13 +72,14 @@ This database powers the **STXBP1-ARIA** therapeutic discovery system, enabling:
 
 ```
 STXBP1-RAG-Database/
-├── chroma.sqlite3                    # Main database (14.9 GB)
-└── d3ded7f6-11aa-4b46-836f-.../     # Index files
-    ├── data_level0.bin               # HNSW index data (2.0 GB)
-    ├── header.bin                    # Index header
-    ├── index_metadata.pickle         # Index metadata
-    ├── length.bin                    # Length data
-    └── link_lists.bin                # HNSW links
+├── chroma.sqlite3                    # Main database
+├── metadata.json                     # Build info
+└── [uuid]/                           # HNSW index files
+    ├── data_level0.bin               # Vector index
+    ├── header.bin                    
+    ├── index_metadata.pickle         
+    ├── length.bin                    
+    └── link_lists.bin                
 ```
 
 ## 🔧 Usage
@@ -75,8 +98,8 @@ db_path = snapshot_download(
     repo_type="dataset"
 )
 
-# Load embedding model (must match indexing model!)
-embedder = SentenceTransformer("all-MiniLM-L6-v2")
+# Load embedding model (MUST match indexing model!)
+embedder = SentenceTransformer("BAAI/bge-base-en-v1.5")
 
 # Connect to ChromaDB
 client = chromadb.PersistentClient(
@@ -88,13 +111,13 @@ client = chromadb.PersistentClient(
 collection = client.get_collection("stxbp1_papers")
 print(f"Loaded {collection.count():,} chunks")
 
-# Search
+# Search (BGE recommends query prefix for retrieval)
 query = "STXBP1 dominant negative mechanism therapeutic approaches"
-query_embedding = embedder.encode([query]).tolist()
+query_embedding = embedder.encode(query, normalize_embeddings=True).tolist()
 
 results = collection.query(
-    query_embeddings=query_embedding,
-    n_results=5,
+    query_embeddings=[query_embedding],
+    n_results=10,
     include=["documents", "metadatas", "distances"]
 )
 
@@ -103,8 +126,8 @@ for doc, meta, dist in zip(
     results['metadatas'][0],
     results['distances'][0]
 ):
-    score = 1 / (1 + dist)  # Convert L2 distance to similarity
-    print(f"[{meta['pmc_id']}] (score: {score:.3f})")
+    pmcid = meta.get('pmcid', meta.get('pmc_id', 'Unknown'))
+    print(f"[{pmcid}] (distance: {dist:.3f})")
     print(f"{doc[:200]}...\n")
 ```
 
@@ -112,78 +135,69 @@ for doc, meta, dist in zip(
 
 See the full retriever implementation at: [STXBP1-Variant-Lookup Space](https://huggingface.co/spaces/SkyWhal3/STXBP1-Variant-Lookup)
 
-## 📚 Source Literature
+## 📚 Curated Corpus
 
-The database indexes PMC papers covering:
+Unlike v8's broad collection, v9 uses a **curated corpus** filtered for STXBP1 relevance:
 
-- **STXBP1/Munc18-1** protein function and mutations
-- **Epileptic encephalopathy** and developmental disorders
-- **Synaptic transmission** mechanisms
-- **Gene therapy** approaches (AAV, base editing, prime editing)
-- **Chemical chaperones** (4-phenylbutyrate, etc.)
-- **Stop codon readthrough** compounds
-- **Protein folding** and aggregation
-- **Clinical trials** and case studies
+### Primary Keywords (Auto-include)
+- STXBP1, Munc18-1, Munc18, syntaxin binding protein
+- UNC-18, N-Sec1
 
-### Search Terms Used for Corpus Collection
+### Related Keywords (Relevance filtered)
+- **Epilepsy**: epileptic encephalopathy, Ohtahara, West syndrome, Dravet, infantile spasms
+- **Synaptic**: SNARE complex, syntaxin-1, synaptic vesicle, exocytosis, neurotransmitter release
+- **Genetics**: haploinsufficiency, dominant negative, nonsense/missense/frameshift mutations
+- **Therapeutics**: gene therapy, AAV, ASO, CRISPR, base editing, prime editing
+- **Chaperones**: 4-PBA, phenylbutyrate, protein folding, proteostasis
+- **Neurodevelopment**: intellectual disability, developmental delay, autism
 
-```
-STXBP1, Munc18-1, syntaxin binding protein, 
-epileptic encephalopathy, developmental epilepsy,
-synaptic vesicle, neurotransmitter release,
-haploinsufficiency, dominant negative,
-chemical chaperone, 4-phenylbutyrate,
-base editing, prime editing, AAV gene therapy,
-stop codon readthrough, ataluren, gentamicin
-```
+### Curated Entries
+Includes 24 hand-curated entries covering:
+- Key primary research (Guiberson 2018, Kovacevic 2018, etc.)
+- Therapeutic mechanism summaries
+- Variant-specific knowledge
+- Clinical trial information
 
 ## 🏗️ How It Was Built
 
-### 1. Paper Collection
-- Queried PubMed Central (PMC) Open Access subset
-- Downloaded full-text HTML/XML for 31,786 papers
-- Extracted text, figures, tables, and captions
+### 1. Corpus Curation
+- Filtered 27,000 multimodal PMC papers by relevance keywords
+- Kept ~17,000 papers with direct STXBP1 relevance
+- Added 41 targeted high-value papers
+- Included 24 curated expert entries
 
-### 2. Text Processing
-- Chunked documents into ~500 token segments
-- Preserved paragraph boundaries where possible
-- Added 50-token overlap between chunks
-- Retained metadata (PMC ID, title, section)
+### 2. Text Processing  
+- Chunked documents into ~1500 character segments
+- 200 character overlap between chunks
+- Preserved document metadata (PMC ID, title)
 
 ### 3. Embedding Generation
-- Used `sentence-transformers/all-MiniLM-L6-v2`
-- 384-dimensional embeddings
-- Batch processing with GPU acceleration
+- Used `BAAI/bge-base-en-v1.5` (768 dimensions)
+- Normalized embeddings for cosine similarity
+- GPU-accelerated batch processing
 
 ### 4. Index Building
 - ChromaDB with persistent storage
-- HNSW (Hierarchical Navigable Small World) index
-- Optimized for semantic similarity search
-
-### Processing Time
-- Indexing: ~5 hours 42 minutes on AMD 5950X
-- Total chunks processed: 1,194,693 (~500 token chunk)
+- HNSW index optimized for semantic search
+- Built on RTX 3080 in ~55 minutes
 
 ## 📋 Metadata Schema
 
-Each chunk includes metadata:
+Each chunk includes:
 
 ```json
 {
-  "pmc_id": "PMC1234567",
+  "pmcid": "PMC1234567",
   "title": "Paper title",
-  "section": "Introduction",
-  "chunk_index": 0,
-  "source_type": "text"
+  "chunk_idx": 0,
+  "source": "multimodal_corpus"
 }
 ```
 
-## ⚠️ Limitations
-
-- **English only** - Non-English papers excluded
-- **PMC Open Access** - Does not include paywalled literature
-- **Static snapshot** - Papers published after indexing date not included
-- **Chunk boundaries** - Some context may be split across chunks
+Source types:
+- `multimodal_corpus` - Papers from curated PMC collection
+- `targeted_paper` - High-priority STXBP1 papers
+- `curated` - Hand-written expert entries
 
 ## 🔬 Use Cases
 
@@ -193,26 +207,31 @@ Each chunk includes metadata:
 4. **Clinical Context** - Find case reports and trial results
 5. **Literature Review** - Rapid survey of research landscape
 
-## 📄 Citation
+## ⚡ Performance Notes
 
-If you use this database in your research, please cite:
+- **Free Tier Compatible**: BGE-base runs on CPU or minimal GPU
+- **Query Time**: <100ms typical retrieval
+- **Memory**: ~1-2GB RAM for embedding model
+
+## 🔗 Related Resources
+
+- **STXBP1-ARIA MAX** (Nemotron embeddings): Coming soon
+- **STXBP1-Variant-Lookup**: [HuggingFace Space](https://huggingface.co/spaces/SkyWhal3/STXBP1-Variant-Lookup)
+- **STXBP1 Foundation**: [stxbp1disorders.org](https://www.stxbp1disorders.org/)
+- **ClinVar STXBP1**: [NCBI ClinVar](https://www.ncbi.nlm.nih.gov/clinvar/?term=STXBP1)
+
+## 📄 Citation
 
 ```bibtex
-@dataset{stxbp1_rag_database_2024,
+@dataset{stxbp1_rag_database_2026,
   author = {Freygang, Adam},
-  title = {STXBP1-ARIA RAG Database: A Vector Index of Biomedical Literature for Therapeutic Discovery},
-  year = {2024},
+  title = {STXBP1-ARIA RAG Database v9: BGE-Embedded Biomedical Literature for Therapeutic Discovery},
+  year = {2026},
   publisher = {HuggingFace},
   url = {https://huggingface.co/datasets/SkyWhal3/STXBP1-RAG-Database}
 }
 ```
 
-## 🔗 Related Resources
-
-- **STXBP1-ARIA Space**: [huggingface.co/spaces/SkyWhal3/STXBP1-Variant-Lookup](https://huggingface.co/spaces/SkyWhal3/STXBP1-Variant-Lookup)
-- **STXBP1 Foundation**: [stxbp1disorders.org](https://www.stxbp1disorders.org/)
-- **ClinVar STXBP1**: [ncbi.nlm.nih.gov/clinvar/?term=STXBP1](https://www.ncbi.nlm.nih.gov/clinvar/?term=STXBP1)
-
 ## 📧 Contact
 
 **Adam Freygang**  
@@ -223,4 +242,4 @@ AI/ML Engineer & STXBP1 Parent Researcher
 
 *Built with ❤️ for the STXBP1 community*
 
-*Part of the NeuroSenpai v3 + STXBP1-ARIA therapeutic discovery system*
+*Part of the NeuroSenpai + STXBP1-ARIA therapeutic discovery system*