NVIDIA-AI-Blueprints · antoniomtz · Apr 17, 2026 · Apr 17, 2026
diff --git a/AGENTS.md b/AGENTS.md
@@ -12,14 +12,18 @@ This document provides guidelines and instructions for AI assistants working on
 - **[README.md](README.md)** - Quick start guide and high-level overview
 - **[docs/API.md](docs/API.md)** - Complete API reference with examples
 - **[docs/DOCKER.md](docs/DOCKER.md)** - Docker and container deployment guide
-- **[PRD.md](PRD.md)** - Product requirements document
+- **[docs/PRD.md](docs/PRD.md)** - Product requirements document
+- **[docs/POLICY_COMPLIANCE.md](docs/POLICY_COMPLIANCE.md)** - Policy compliance feature guide
+- **[docs/PRODUCT_MANUAL_FAQS.md](docs/PRODUCT_MANUAL_FAQS.md)** - Product manual PDF for FAQ enrichment guide
 - **[AGENTS.md](AGENTS.md)** - This file (AI assistant guidelines)
 
 ### Current Status
 - ✅ **Multi-Language Support** - Locale-based product descriptions (FR-6 completed)
 - ✅ **VLM Content Augmentation** - Enhances existing product data with visual insights (FR-2 completed)
 - ✅ **2D Image Variation Generation** - Working with prompt planning and quality evaluation (FR-3 completed)
 - ✅ **Automated Quality Assessment** - VLM-based reflection for generated images (FR-9 completed)
+- ✅ **Product FAQ Generation** - FAQs from enriched data with optional product manual PDF enhancement (FR-10, FR-12 completed)
+- ✅ **Policy Compliance** - PDF policy library with Milvus RAG and compliance classification (FR-11 completed)
 - ⚠️ **In Development** - 3D Asset Generation (backend complete) and Video Generation in progress
 
 ### Key Goals
@@ -44,7 +48,7 @@ cd catalog-enrichment
 ### Backend (current)
 
 - Stack: FastAPI + Uvicorn (ASGI), OpenAI client (NVIDIA endpoint), Starlette under the hood
-- Dependencies: `fastapi`, `uvicorn[standard]`, `openai`, `python-multipart`, `python-dotenv`, `httpx`, `pillow`, `pyyaml`
+- Dependencies: `fastapi`, `uvicorn[standard]`, `openai`, `python-multipart`, `python-dotenv`, `httpx`, `pillow`, `pyyaml`, `pymilvus`, `pypdf`, `numpy`
 - Python: 3.11+
 - **Error Handling**: Comprehensive connection error detection with user-friendly messages when NIM endpoints are unreachable
 
@@ -102,6 +106,32 @@ uvicorn --app-dir src backend.main:app --host 0.0.0.0 --port 8000 --reload
     - `quality_score`: float (0-100 quality score from VLM reflection, or null if evaluation failed)
     - `quality_issues`: array (list of detected quality issues from reflection analysis)
 
+**FAQ Generation:**
+- POST `/vlm/faqs`
+  - Request: `multipart/form-data` with fields:
+    - `title` (string, optional): Product title from VLM analysis
+    - `description` (string, optional): Product description from VLM analysis
+    - `categories` (JSON string, optional): Categories array
+    - `tags` (JSON string, optional): Tags array
+    - `colors` (JSON string, optional): Colors array
+    - `locale` (string, optional): Regional locale code (default: "en-US")
+    - `manual_knowledge` (JSON string, optional): Extracted knowledge from `/vlm/manual/extract`
+  - Response: `{ "faqs": [{ "question": "string", "answer": "string" }] }`
+  - Without manual: 3-5 FAQs from product data
+  - With manual knowledge: up to 10 FAQs drawing from both product data and manual
+
+**Product Manual Knowledge Extraction:**
+- POST `/vlm/manual/extract`
+  - Request: `multipart/form-data` with fields:
+    - `file` (file): Product manual PDF (max 50 MB)
+    - `title` (string, optional): Product title for query generation
+    - `categories` (JSON string, optional): Product categories for query generation
+    - `locale` (string, optional): Regional locale code (default: "en-US")
+  - Response: `{ "filename": "string", "chunk_count": 42, "knowledge": { "topic": "extracted text..." } }`
+  - Stateless: all vectors freed after response, no server-side storage
+  - LLM generates 5-8 product-type-specific queries from title + categories (not description)
+  - Retrieves relevant chunks per query via in-memory cosine similarity
+
 **3D Asset Generation:**
 - POST `/generate/3d`
   - Request: `multipart/form-data` with fields:
@@ -344,7 +374,7 @@ Given the catalog enrichment focus, pay special attention to:
 
 ---
 
-**Last Updated:** $(date)  
-**Version:** 1.0  
+**Last Updated:** 16-Apr-2026  
+**Version:** 1.3  
 
 *This document should be updated as the project evolves and new practices are established.*
diff --git a/README.md b/README.md
@@ -25,15 +25,17 @@ A GenAI-powered catalog enrichment system that transforms basic product images i
 - **Cultural Image Generation**: Create culturally-appropriate product backgrounds (Spanish courtyards, Mexican family spaces, British formal settings)
 - **Quality Evaluation**: Automated VLM-based quality assessment of generated images with detailed scoring
 - **3D Asset Generation**: Transform 2D product images into interactive 3D GLB models using Microsoft TRELLIS
-- **Product FAQ Generation**: Automatically generate 3-5 product FAQs from enriched catalog data
+- **Product FAQ Generation**: Automatically generate product FAQs from enriched catalog data, with optional product manual PDF upload for richer FAQs (up to 10) via stateless targeted RAG
 - **Policy Compliance**: Upload policy PDFs and automatically check product listings against them using RAG + Milvus
 - **Modular API**: Separate endpoints for VLM analysis, FAQ generation, image generation, and 3D asset generation
 
 ## Documentation
 
 - **[API Documentation](docs/API.md)** - Detailed API endpoints, parameters, and examples
 - **[Docker Deployment Guide](docs/DOCKER.md)** - Docker and Docker Compose setup instructions
-- **[Product Requirements (PRD)](PRD.md)** - Product requirements and feature specifications
+- **[Product Requirements (PRD)](docs/PRD.md)** - Product requirements and feature specifications
+- **[Policy Compliance](docs/POLICY_COMPLIANCE.md)** - How policy compliance checking works
+- **[Product Manual for FAQs](docs/PRODUCT_MANUAL_FAQS.md)** - How product manual PDFs enrich FAQ generation
 - **[AI Agent Guidelines](AGENTS.md)** - Instructions for AI assistants working on this project
 
 ## Tech Stack
@@ -221,6 +223,8 @@ For complete Docker deployment instructions, see the **[Docker Deployment Guide]
 The system provides three main endpoints:
 
 - `POST /vlm/analyze` - Fast VLM/LLM analysis
+- `POST /vlm/faqs` - Product FAQ generation (supports optional manual knowledge)
+- `POST /vlm/manual/extract` - Extract knowledge from a product manual PDF for FAQ enrichment
 - `POST /generate/variation` - Image generation with FLUX
 - `POST /generate/3d` - 3D asset generation with TRELLIS
 

diff --git a/docs/API.md b/docs/API.md
@@ -37,6 +37,7 @@ The API provides a modular approach for optimal performance and flexibility:
 
 **1) Fast VLM Analysis (POST `/vlm/analyze`)** - Get product fields quickly
 **2) FAQ Generation (POST `/vlm/faqs`)** - Generate product FAQs from enriched data
+**2.5) Manual Knowledge Extraction (POST `/vlm/manual/extract`)** - Extract knowledge from a product manual PDF to enrich FAQs
 **3) Image Generation (POST `/generate/variation`)** - Generate 2D variations on demand  
 **4) 3D Asset Generation (POST `/generate/3d`)** - Generate 3D models on demand
 
@@ -276,7 +277,10 @@ curl -X POST \
 
 ## 3️⃣ FAQ Generation: `/vlm/faqs`
 
-Generate 3-5 frequently asked questions and answers for a product based on its enriched catalog data. Designed to be called after `/vlm/analyze` completes, using the enriched result as input.
+Generate frequently asked questions and answers for a product based on its enriched catalog data. Designed to be called after `/vlm/analyze` completes, using the enriched result as input.
+
+Without a product manual: generates 3-5 basic FAQs from the product data.
+With manual knowledge (from `/vlm/manual/extract`): generates up to 10 richer FAQs that draw from both the product data and the manual, surfacing details that go beyond the description.
 
 **Endpoint**: `POST /vlm/faqs`
 **Content-Type**: `multipart/form-data`
@@ -291,6 +295,7 @@ Generate 3-5 frequently asked questions and answers for a product based on its e
 | `tags` | JSON string | No | Tags array (default: `[]`) |
 | `colors` | JSON string | No | Colors array (default: `[]`) |
 | `locale` | string | No | Regional locale code (default: `en-US`) |
+| `manual_knowledge` | JSON string | No | Extracted manual knowledge from `/vlm/manual/extract` |
 
 ### Response Schema
 
@@ -305,7 +310,7 @@ Generate 3-5 frequently asked questions and answers for a product based on its e
 }
 ```
 
-### Usage Example
+### Usage Example (Basic)
 
 ```bash
 # Call after /vlm/analyze to generate FAQs from enriched data
@@ -319,6 +324,27 @@ curl -X POST \
   http://localhost:8000/vlm/faqs
 ```
 
+### Usage Example (With Product Manual)
+
+```bash
+# First extract knowledge from the manual, then pass it to FAQ generation
+KNOWLEDGE=$(curl -s -X POST \
+  -F "file=@mower-manual.pdf" \
+  -F "title=Craftsman 20V Cordless Lawn Mower" \
+  -F 'categories=["electronics"]' \
+  http://localhost:8000/vlm/manual/extract | jq -c '.knowledge')
+
+curl -X POST \
+  -F "title=Craftsman 20V Cordless Lawn Mower" \
+  -F "description=A cordless lawn mower featuring a black and red design..." \
+  -F 'categories=["electronics"]' \
+  -F 'tags=["cordless","lawn mower","Craftsman"]' \
+  -F 'colors=["black","red"]' \
+  -F "locale=en-US" \
+  -F "manual_knowledge=$KNOWLEDGE" \
+  http://localhost:8000/vlm/faqs
+```
+
 ### Example Response
 
 ```json
@@ -342,6 +368,77 @@ curl -X POST \
 
 ---
 
+## 3.5️⃣ Product Manual Knowledge Extraction: `/vlm/manual/extract`
+
+Extract structured knowledge from a product manual PDF using targeted RAG. The endpoint processes the PDF, generates product-type-specific queries via the LLM (using title + categories, not description, to avoid duplicating what the description already covers), and retrieves relevant chunks from the manual for each topic.
+
+This endpoint is **stateless** — all embeddings are computed in-memory and freed after the response. It can handle concurrent requests for different products.
+
+**Endpoint**: `POST /vlm/manual/extract`
+**Content-Type**: `multipart/form-data`
+
+### Request Parameters
+
+| Parameter | Type | Required | Description |
+|-----------|------|----------|-------------|
+| `file` | file | Yes | Product manual PDF (max 50 MB) |
+| `title` | string | No | Product title (used to generate relevant queries) |
+| `categories` | JSON string | No | Product categories array (used to generate relevant queries) |
+| `locale` | string | No | Regional locale code (default: `en-US`) |
+
+### Response Schema
+
+```json
+{
+  "filename": "string",
+  "chunk_count": 42,
+  "knowledge": {
+    "battery_life": "The speaker provides up to 12 hours of continuous playback...",
+    "waterproof_rating": "IPX7 rated, can be submerged up to 1 meter for 30 minutes...",
+    "care_instructions": "Clean with a damp cloth. Do not use abrasive cleaners..."
+  }
+}
+```
+
+The `knowledge` object contains topic keys (dynamically generated by the LLM based on product type) mapped to the relevant text extracted from the manual. Topics with no relevant content are empty strings.
+
+### Usage Example
+
+```bash
+curl -X POST \
+  -F "file=@speaker-manual.pdf;type=application/pdf" \
+  -F "title=JBL Flip 6 Portable Speaker" \
+  -F 'categories=["electronics"]' \
+  -F "locale=en-US" \
+  http://localhost:8000/vlm/manual/extract
+```
+
+### Batch Script Example
+
+```bash
+# Process multiple products concurrently (each request is independent)
+for product in products/*.json; do
+  TITLE=$(jq -r '.title' "$product")
+  CATS=$(jq -c '.categories' "$product")
+  PDF=$(jq -r '.manual_pdf' "$product")
+
+  KNOWLEDGE=$(curl -s -X POST \
+    -F "file=@$PDF" \
+    -F "title=$TITLE" \
+    -F "categories=$CATS" \
+    http://localhost:8000/vlm/manual/extract | jq -c '.knowledge')
+
+  curl -s -X POST \
+    -F "title=$TITLE" \
+    -F "description=$(jq -r '.description' "$product")" \
+    -F "categories=$CATS" \
+    -F "manual_knowledge=$KNOWLEDGE" \
+    http://localhost:8000/vlm/faqs
+done
+```
+
+---
+
 ## 4️⃣ Image Generation: `/generate/variation`
 
 Generate culturally-appropriate product variations using FLUX models based on VLM analysis results.