Initial commit: RAG demo with build and query scripts

.gitignore

@@ -0,0 +1,15 @@
+# Python
+.venv/
+__pycache__/
+*.pyc
+
+# Generated data
+storage/
+models/
+
+# macOS
+.DS_Store
+
+# Data (users supply their own)
+data/*
+!data/.gitkeep

README.md

@@ -0,0 +1,36 @@
+# RAG Demo
+
+Retrieval Augmented Generation using LlamaIndex with local models.
+
+This demo builds a semantic search system over a collection of text documents
+using a HuggingFace embedding model and Ollama for generation.
+
+## Tutorial
+
+See the full walkthrough at:
+https://lem.che.udel.edu/wiki/index.php?n=Main.RAG
+
+## Quick Start
+
+```bash
+# Create and activate virtual environment
+python3 -m venv .venv
+source .venv/bin/activate
+
+# Install dependencies
+pip install -r requirements.txt
+
+# Pull the generating model
+ollama pull command-r7b
+
+# Place your .txt documents in ./data, then build the vector store
+python build.py
+
+# Run interactive queries
+python query.py
+```
+
+## Models
+
+- **Embedding:** BAAI/bge-large-en-v1.5 (downloaded automatically on first run)
+- **Generation:** command-r7b via Ollama

build.py

@@ -0,0 +1,49 @@
+# build.py
+#
+# Import documents from data, generate embedded vector store
+# and save to disk in directory ./storage
+#
+# August 2025
+# E. M. Furst
+
+from llama_index.core import (
+    SimpleDirectoryReader,
+    VectorStoreIndex,
+    Settings,
+)
+
+from llama_index.embeddings.huggingface import HuggingFaceEmbedding
+from llama_index.core.node_parser import SentenceSplitter
+
+def main():
+    # Choose your embedding model
+    embed_model = HuggingFaceEmbedding(cache_folder="./models",
+        model_name="BAAI/bge-large-en-v1.5")
+
+    # Configure global settings for LlamaIndex
+    Settings.embed_model = embed_model
+
+    # Load documents
+    documents = SimpleDirectoryReader("./data").load_data()
+
+    # Create the custom textsplitter
+    # Set chunk size and overlap (e.g., 256 tokens, 25 tokens overlap)
+    text_splitter = SentenceSplitter(
+        chunk_size=500,
+        chunk_overlap=50,
+    )
+    Settings.text_splitter = text_splitter
+
+    # Build the index
+    index = VectorStoreIndex.from_documents(
+        documents, transformations=[text_splitter],
+        show_progress=True,
+    )
+
+    # Persist both vector store and index metadata
+    index.storage_context.persist(persist_dir="./storage")
+
+    print("Index built and saved to ./storage")
+
+if __name__ == "__main__":
+    main()

query.py

@@ -0,0 +1,110 @@
+# query.py
+#
+# Run a query on a vector store
+#
+# August 2025
+# E. M. Furst
+
+from llama_index.core import (
+    load_index_from_storage,
+    StorageContext,
+    Settings,
+)
+from llama_index.embeddings.huggingface import HuggingFaceEmbedding
+from llama_index.llms.ollama import Ollama
+from llama_index.core.prompts import PromptTemplate
+import os, time
+
+#
+# Globals
+#
+os.environ["TOKENIZERS_PARALLELISM"] = "false"
+
+# Embedding model used in vector store (this should match the one in build.py)
+embed_model = HuggingFaceEmbedding(cache_folder="./models",
+    model_name="BAAI/bge-large-en-v1.5")
+
+# LLM model to use in query transform and generation
+llm = "command-r7b"
+
+#
+# Custom prompt for the query engine
+#
+PROMPT = PromptTemplate(
+"""You are an expert research assistant. You are given top-ranked writing \
+excerpts (CONTEXT) and a user's QUERY.
+
+Instructions:
+- Base your response *only* on the CONTEXT.
+- The snippets are ordered from most to least relevant—prioritize insights \
+from earlier (higher-ranked) snippets.
+- Aim to reference *as many distinct* relevant files as possible (up to 10).
+- Do not invent or generalize; refer to specific passages or facts only.
+- If a passage only loosely matches, deprioritize it.
+
+Format your answer in two parts:
+
+1. **Summary Theme**
+   Summarize the dominant theme from the relevant context in a few sentences.
+
+2. **Matching Files**
+   Make a list of 10 matching files. The format for each should be:
+   <filename> - <rationale tied to content. Include date if available.>
+
+CONTEXT:
+{context_str}
+
+QUERY:
+{query_str}
+
+Now provide the theme and list of matching files."""
+)
+
+#
+# Main program routine
+#
+
+def main():
+    # Use a local model to generate -- in this case using Ollama
+    Settings.llm = Ollama(
+        model=llm,
+        request_timeout=360.0,
+    )
+
+    # Load embedding model (same as used for vector store)
+    Settings.embed_model = embed_model
+
+    # Load persisted vector store + metadata
+    storage_context = StorageContext.from_defaults(persist_dir="./storage")
+    index = load_index_from_storage(storage_context)
+
+    # Build regular query engine with custom prompt
+    query_engine = index.as_query_engine(
+        similarity_top_k=15,
+        text_qa_template=PROMPT,
+    )
+
+    # Query
+    while True:
+        q = input("\nEnter a search topic or question (or 'exit'): ").strip()
+        if q.lower() in ("exit", "quit"):
+            break
+        print()
+
+        # Generate the response by querying the engine
+        start_time = time.time()
+        response = query_engine.query(q)
+        end_time = time.time()
+
+        # Return the query response and source documents
+        print(response.response)
+
+        print("\nSource documents:")
+        for node in response.source_nodes:
+            meta = getattr(node, "metadata", None) or node.node.metadata
+            print(f"  {meta.get('file_name')}  {getattr(node, 'score', None)}")
+
+        print(f"\nElapsed time: {(end_time-start_time):.1f} seconds")
+
+if __name__ == "__main__":
+    main()

requirements.txt

@@ -0,0 +1,4 @@
+llama-index-core
+llama-index-readers-file
+llama-index-llms-ollama
+llama-index-embeddings-huggingface