AI Research & Knowledge Assistant

A production-oriented Retrieval-Augmented Generation (RAG) backend application that enables users to upload technical research PDFs, automatically classify them using a Machine Learning predictor, query them using hybrid keyword/semantic search, run session-based Q&A with proper source citations, and generate comparative summaries.

---

1. Architecture Diagram

graph TD
    Client[Client / REST API] --> |Upload PDF| DocRoute[Document API]
    DocRoute --> |Save PDF to Uploads| FS[Filesystem]
    DocRoute --> |Add Record: PENDING| SQL[SQLite DB]
    DocRoute --> |Enqueue Ingestion Task| BG[FastAPI Background Workers]
    
    BG --> |1. Extract Text page-by-page| Parser[PyMuPDF Parser]
    BG --> |2. Consensus Classification| Predictor[ML Predictor]
    Predictor --> |Multinomial NB fallback / TensorFlow NN| Predictor
    Predictor --> |Update Category| SQL
    
    BG --> |3. Chunk with page reference| Chunker[Page Chunker]
    BG --> |4. Index vector embeddings| VectorDB[(Chroma DB Collection)]
    
    Client --> |Search / QA query| SearchRoute[Search API]
    SearchRoute --> |Semantic search| VectorDB
    SearchRoute --> |BM25 Keyword search| VectorDB
    SearchRoute --> |Reciprocal Rank Fusion| RRF[Hybrid RRF Merger]
    RRF --> |Consolidated Context| QA[RAG QA Engine]
    QA --> |Query memory history| SQL
    QA --> |Format Answer with page citations| Client

Loading

---

2. Technology Stack

• Framework: FastAPI (high-performance asynchronous Python REST framework)
• Vector Store: Chroma DB (embedding database for semantic storage and retrieval)
• Database: SQLite (SQL engine managed via SQLAlchemy ORM for metadata, metrics, and chat history)
• Document Parsing: PyMuPDF (extremely fast layout-preserving PDF parser)
• Embedding Model: all-MiniLM-L6-v2 via HuggingFace Hub (downloaded locally by Chroma)
• LLM Provider: LangChain / OpenAI API (gpt-4o-mini / gpt-3.5-turbo)
• Machine Learning: TensorFlow 2.x (Neural Network classification) + Scikit-Learn (Multinomial Naive Bayes pipeline fallback)

---

3. Setup Instructions

Prerequisites

• Python 3.10 to 3.13
• Git

Installation

1. Clone the Repository:

   git clone https://github.com/ananyadarna/AI-Research-Knowledge-Assistant.git
   cd "AI-Research-Knowledge-Assistant"

2. Create and Activate Virtual Environment:

   python -m venv .venv
   # Windows
   .venv\Scripts\activate
   # macOS/Linux
   source .venv/bin/activate

3. Install Pinned Dependencies:

   pip install -r requirements.txt

4. Configure Environment Variables: Create a .env file from the template:

   copy .env.example .env

   Open .env and fill in your API key (e.g. OPENAI_API_KEY=your_key).

5. Pre-train Classification Assets (Optional but Recommended): Train the document classifier locally on synthetic dataset:

   python -m src.ml.train_classifier

6. Run FastAPI Application:

   python main.py
   # Server will start on http://127.0.0.1:8000

7. Run Unit & Integration Test Suite:

   python -m pytest -v

Deployment

• Platform: Render Cloud (Render Web Service)
• Configuration: render.yaml
• Live Production URL: https://ai-research-knowledge-assistant-erlk.onrender.com
• Live Interactive Swagger Docs: https://ai-research-knowledge-assistant-erlk.onrender.com/docs
• Live ReDoc Specifications: https://ai-research-knowledge-assistant-erlk.onrender.com/redoc

---

4. Environment Variables

Define the following in your .env file:

Variable	Description	Default / Example
OPENAI_API_KEY	OpenAI API credential for RAG, summary, and comparisons	sk-...
EMBEDDING_PROVIDER	Provider for text vectorization (openai or local)	local
DATABASE_URL	SQLAlchemy connection string	sqlite:///./data/assistant.db
VECTOR_DB_DIR	Chroma DB persistence storage directory	./data/vector_db
UPLOAD_DIR	Location to save uploaded source PDF files	./data/uploads
MODEL_PATH	Path where trained TF classification models reside	./models/tf_classifier.h5
SKLEARN_MODEL_PATH	Path to the scikit-learn Naive Bayes fallback model	./models/sklearn_classifier.pickle

---

5. API Documentation

For full Swagger interactive docs, navigate to http://127.0.0.1:8000/docs while the server is running.

5.1 Document Management

• Upload Document: POST /documents/upload
  • Payload: Multipart form file (file: UploadFile)
  • Response: {"doc_id": "...", "processing_status": "PENDING"}
• List Documents: GET /documents
  • Response: Returns a JSON list of document metadata.
• Reprocess Document: POST /documents/{doc_id}/reprocess
  • Response: Wipes vector store elements and resets processing state to PENDING.
• Delete Document: DELETE /documents/{doc_id}
  • Response: Wipes SQL record, vector DB chunks, and filesystem upload.

5.2 Search & RAG Q&A

• Search Retrieval: POST /search/query
  • Payload: {"query": "neural networks", "search_mode": "hybrid", "doc_ids": ["uuid"], "top_k": 3}
  • Response: List of ranked chunks with relevance scores.
• RAG QA Conversation: POST /search/qa
  • Payload: {"session_id": "optional-session", "query": "What is the training accuracy?", "doc_ids": ["uuid"]}
  • Response: Answer text + citations list (containing source text, filename, and page numbers).

5.3 Summary & Comparison

• Document Summarization: POST /analysis/summarize/{doc_id}
  • Response: JSON block containing executive_summary, technical_summary, bullet_points, and key_takeaways.
• Multi-Document Comparison: POST /analysis/compare
  • Payload: {"doc_ids": ["uuid1", "uuid2"]}
  • Response: Comparative grid outlining methodologies, advantages, disadvantages, differences, and conclusions.

5.4 Analytics Dashboard

• System Telemetry Stats: GET /analytics/stats
  • Response: JSON containing metrics (total_documents, total_chunks, total_embeddings_generated, total_questions_answered, category_distribution, most_queried_documents).

---

6. Assumptions & Design Decisions

6.1 Page-by-Page Chunking Strategy

Instead of splitting text purely by character counts which overlaps pages randomly, we partition text page-by-page. This design decision guarantees that:

1. Citations are 100% accurate: Each chunk corresponds to precisely one physical PDF page number.
2. No context boundary bleed: Avoids mixing text context from neighboring unrelated pages.

6.2 Reciprocal Rank Fusion (RRF) Hybrid Search

Combining keyword matches (BM25) and semantic vector distance requires joining different scoring bounds. We implemented RRF: $$\text{Score}(d) = \sum_{m \in M} \frac{1}{60 + \text{Rank}_m(d)}$$ This ranks chunks that have high scores in both modes higher, providing maximum search precision.

6.3 Dual-Model Machine Learning Fallback

• TensorFlow Neural Net: Multi-layer Dense Net with OOV Embedding layers.
• Scikit-Learn NB Fallback: Uses TfidfVectorizer + MultinomialNB. If a Python 3.13 workspace or environment has network limits blocking TensorFlow downloads, the system automatically falls back to loading scikit-learn models.

6.4 Resilient Vector Store Manager Fallback

• Chroma DB: Uses Chroma DB persistent collection by default.
• SQLite Vector Store Fallback: If Windows Application Control (AppLocker) or python virtual environments restrict loading third-party C-DLL binaries (cygrpc.pyd), VectorSearchManager automatically initializes a pure-python/SQLite persistent vector store fallback with zero startup downtime.

---

7. Limitations & Future Improvements

1. Symlink Support on Windows: Windows systems without Developer Mode enabled may show warning logs when Chroma caches HuggingFace tokenizers. Standard file copies fallback automatically.
2. OCR Support: Currently assumes text is readable. Scanned document indexing would require Tesseract/OCR libraries.
3. Reranking Models: Adding Cohere or Cross-Encoder model rerankers after RRF would further improve citation scoring in multi-document comparisons.