docs/ANSWER.md

Certification Challenge

Marko Budisic

Deliverables

1. Main Github repo
2. Github repo for creating the golden dataset
3. Written document
4. Loom video
5. Hugging Face live demo
6. Fine tuned embedding model
7. Corpus dataset
8. Golden Q&A dataset

ToC

• Certification Challenge
  • Deliverables
  • ToC
  • Task 1: Defining your Problem and Audience
  • Task 2: Propose a Solution
  • Task 3: Dealing with the Data
    • 3.1. Data Sources & External APIs 📊+🌐
    • 3.2. Chunking Strategy 🧠✂️
    • 3.3. Specific Data Needs for Other Parts 🧩
  • Task 4: Building a Quick End-to-End Prototype
    • 4.1. The Prototype Application 🖥️
    • 4.2. Deployment 🚀 (Hugging Face Space)
  • Task 5: Creating a Golden Data Set
    • 5.1. RAGAS Framework Assessment & Results 📊
  • Task 6: Fine-Tuning Open-Source Embeddings
    • 6.1. Fine-Tuning Process and Model Link 🔗:**
  • Task 7: Assessing Performance
    • 7.1. Comparative RAGAS Results 📊
  • 8. Future changes

Task 1: Defining your Problem and Audience

Problem: In large orgs with legacy video archives, navigating extensive libraries of video materials to find specific information and build reports and materials based on them is a time-consuming and inefficient process. 😓

Users and their Problem: 🏢 Companies often have extensive video tutorial libraries for proprietary software. Employees (new hires, support, experienced users) struggle to quickly find specific instructions within these videos. 🎯 Like Photoshop learners needing a specific technique, employees need a fast way to query video content, saving time and boosting learning. 🚀

Side note: This is a good approximation of a problem that I am internally solving for my company. The agentic RAG will be augmented further for the demo day.

Task 2: Propose a Solution

Our Solution: 🗣️ An agentic Retrieval Augmented Generation (RAG) system designed to answer questions about a company's video tutorial library (e.g., for software like Adobe Photoshop, or any internal training content). Users interact via a chat interface (Chainlit, as seen in app.py). 💻 The system queries its knowledge base of tutorial transcripts and can use Tavily for web searches, providing comprehensive answers relevant to the specific video library and serving up videos at the referenced timestampes. 🌐

Broader vision is to build an ingestion pipeline that would transcribe audio narration and OCR key frames in the video to further enhance the context. Users would be able to search not only by a query, but also by a screenshot (e.g. looking up live video if they have only a screenshot in a company walkthrough). The agents would not only be able to answer the queries, but also develop a short presentation, e.g., in reveal.js or remark.

The Tech Stack 🛠️:

• LLM: OpenAI model (gpt-4.1-mini), selected for strong language capabilities and ease of API access. 🧠
• Embedding Model: An open-source model, Snowflake/snowflake-arctic-embed-s (see Fine_Tuning_Embedding_for_PSTuts.ipynb), fine-tuned for domain-specific relevance. This is a small model trainable on a laptop. ❄️
• Orchestration: LangChain & LangGraph, for building the RAG application and managing agent workflows. Many functions have been stored in the pstuts_rag package to allow calling from notebooks and app. 🔗
• Vector Database: Qdrant (pstuts_rag/datastore.py), for efficient semantic search of tutorial transcripts. I had most experience with it, and no reason to look elsewhere. 💾
• Evaluation: Synthetic data set, created using RAGAS in a second repo, powers evaluate_rag.ipynb, for assessing RAG pipeline (w/o the search powers) quality. 🧐
• Monitoring: W&B (Weights & Biases)🏋️ was used to monitor fine-tuning. LangSmith was enabled for monitoring in general.📊
• User Interface: Chainlit chat with on-demand display of videos positioned at the correct timestamp. 💬 📼
• Serving & Inference: Docker (Dockerfile), for containerized deployment on Hugging Face Spaces. 🐳

The Role of Agents 🕵️‍♂️:

The system uses a LangGraph-orchestrated multi-agent approach:

1. Supervisor Agent: Manages the overall workflow. It receives the user query and routes it to the appropriate specialized agent based on its interpretation of the query (defined in SUPERVISOR_SYSTEM prompt and create_team_supervisor in app.py). 🧑‍✈️
2. Video Archive Agent (VIDEOARCHIVE): This is the RAG agent. It queries the Qdrant vector store of Photoshop tutorial transcripts to find relevant information and generates an answer based on this retrieved context. (Uses create_rag_node from pstuts_rag.agent_rag). 📼
3. Adobe Help Agent (ADOBEHELP): This agent uses the Tavily API to perform web searches if the supervisor deems it necessary for broader or more current information. (Uses create_tavily_node from pstuts_rag.agent_tavily). 🌍 The supervisor then determines if the answer is complete or if further steps are needed. ✅

                            +-----------+                             
                            | __start__ |                             
                            +-----------+                             
                                  *                                   
                                  *                                   
                                  *                                   
                            +------------+                            
                            | supervisor |                            
                       *****+------------+.....                       
                   ****            .           ....                   
              *****                .               .....              
           ***                     .                    ...           
+-----------+           +--------------------+           +---------+  
| AdobeHelp |           | VideoArchiveSearch |           | __end__ |  
+-----------+           +--------------------+           +---------+  

Task 3: Dealing with the Data

3.1. Data Sources & External APIs 📊+🌐

• Primary Data: PsTuts-VQA is a publicly-released set of transcripts linked to a database of Adobe-created Photoshop training videos. Data is in a JSON format, made available on hf.co:mbudisic/PsTuts-VQA. 📁
• External API: Tavily Search API (configured in app.py) augments knowledge with web search results of domain helpx.adobe.com via the ADOBEHELP agent for current or broader topics not covered in the internal videos. 🔍

3.2. Chunking Strategy 🧠✂️

(see: pstuts_rag/datastore.py's chunk_transcripts function and pstuts_rag/loader.py)

Transcript chunks in the input dataset are too granular - often a sentence or two, since they are tied to the time windows in which a particular transcript sentence would be overlaid on the screen.

Therefore, to achieve a useful semantic chunking for RAG, the following semantic chunking strategy is employed:

1. Initial Loading: Transcripts are loaded both entirely per video (VideoTranscriptBulkLoader) and as individual sentences/segments with timestamps (VideoTranscriptChunkLoader).
2. Semantic Splitting: SemanticChunker (LangChain, using OpenAIEmbeddings) splits full transcripts into semantically coherent chunks.
3. Metadata Enrichment: These semantic chunks are enriched with start/end times by mapping them back to the original timestamped sentences.

In summary: 🤔 This method (a) creates topically focused chunks for better retrieval. 🎯 (b) links back to video timestamps. 🔗

3.3. Specific Data Needs for Other Parts 🧩

• Evaluation & Golden Dataset (Tasks 5 & 7): 🏆 Generating the "Golden Data Set" using Knowledge Graph to produce question-answer-context triplet in RAGAS is detailed in create_golden_dataset.ipynb (see PsTuts-VQA-Data-Operations repo). The resulting dataset hf.co:mbudisic/pstuts_rag_-_qa is used to benchmark the RAG pipeline in evaluate_rag.ipynb and fine-tune the embedding model. 📊

• Embedding Model Fine-Tuning (Task 6): 🔬 The Fine_Tuning_Embedding_for_PSTuts.ipynb notebook shows the use of hf.co:mbudisic/pstuts_rag_qa to fine-tune the embedding model. This adapts models like Snowflake/snowflake-arctic-embed-s for improved retrieval. ⚙️

Task 4: Building a Quick End-to-End Prototype

An end-to-end prototype RAG system for Photoshop tutorials is built and deployed to HF.

4.1. The Prototype Application 🖥️

The app.py script is the core prototype. It uses Chainlit for the UI, LangChain/LangGraph for orchestration, Qdrant for the vector store, and OpenAI models for embeddings and generation. It loads data, builds the RAG chain, and manages the agentic workflow for user queries. ✨

4.2. Deployment 🚀 (Hugging Face Space)

The repository is structured for Hugging Face Space deployment:

• README.md contains Hugging Face Space metadata.
• Dockerfile enables containerization for deployment. Note the load of web dependencies from pyproject.toml.

Task 5: Creating a Golden Data Set

The creation of the "Golden Data Set" is documented in the create_golden_dataset.ipynb notebook in the PsTuts-VQA-Data-Operations repository. This dataset was then utilized in the notebooks/evaluate_rag.ipynb of the current project to assess the initial RAG pipeline with RAGAS. 🌟

5.1. RAGAS Framework Assessment & Results 📊

The initial RAG pipeline ("Base" model, Snowflake/snowflake-arctic-embed-s before fine-tuning) yielded these mean RAGAS scores:

Metric	Mean Score
Faithfulness	0.721
Answer Relevancy	0.914
Context Recall	0.672
Factual Correctness (mode=f1)	0.654
Context Entity Recall	0.636

(Scores from notebooks/evaluate_rag.ipynb output for the "Base" configuration)

2. Conclusions on Performance and Effectiveness 🧐:

• Strengths: 💪 High Answer Relevancy (0.914) indicates the system understands queries well.
• Areas for Improvement: 📉
  • Faithfulness (0.721): Answers are not always perfectly grounded in retrieved context. Maybe if I turned the temperature down to 0 this score would have been higher.
  • Context Recall (0.672): Not all necessary information is always retrieved.
  • Factual Correctness (0.654): Factual accuracy of answers needs improvement.
• Overall: The baseline system is good at relevant responses but needs better context retrieval and factual accuracy. This benchmarks a clear path for improvements, such as embedding fine-tuning. 🛠️

Task 6: Fine-Tuning Open-Source Embeddings

To enhance retrieval performance for a specific video library, an open-source embedding model can be fine-tuned on domain-specific data. The following describes an example of this process using Photoshop tutorial data. 🧪

6.1. Fine-Tuning Process and Model Link 🔗:**

• Base Model: Snowflake/snowflake-arctic-embed-s was chosen as the base model for fine-tuning in this example. The -s stands for small -- larger two models ended up taking too much GPU memory on my laptop. ❄️
• Fine-tuning Data: The fine-tuning notebook is notebooks/Fine_Tuning_Embedding_for_PSTuts.ipynb. It uses the golden dataset, retrieved from the HF repository. 🖼️ The data was split into train-validate-test blocks. train was used to compute the objective function in the training loop, while validate was used in evaluation.
• Monitoring: 🛠️ W&B tracked the process and evaluation. 📈
• Resulting Model: The fine-tuned model (for the Photoshop example) was saved and pushed to the Hugging Face Hub. 🤗 mbudisic/snowflake-arctic-embed-s-ft-pstuts

(See notebooks/Fine_Tuning_Embedding_for_PSTuts.ipynb, specifically the model.push_to_hub call and its output. The app.py is configured to use such a fine-tuned model for the embedding step in the RAG pipeline.)

Task 7: Assessing Performance

Performance of the RAG application with the fine-tuned embedding model (mbudisic/snowflake-arctic-embed-s-ft-pstuts) was assessed using the same RAGAS framework and "Golden Data Set" (golden_small_hf) as the baseline. 🏆

7.1. Comparative RAGAS Results 📊

(see: notebooks/evaluate_rag.ipynb output)

The notebook provides a comparison between "Base", "SOTA" (OpenAI's text-embedding-3-small), and "FT" (our fine-tuned mbudisic/snowflake-arctic-embed-s-ft-pstuts) models.

Metric	Base (Initial)	FT (Fine-Tuned)	Change FT vs Base
Faithfulness	0.721	0.748	+0.027
Answer Relevancy	0.914	0.819	-0.095
Context Recall	0.672	0.672	0.000
Factual Correctness	0.654	0.598	-0.056
Context Entity Recall	0.636	0.636	0.000

(Note: These are mean scores.)

Additionally, statistical significance of changes Base -> FT and FT -> SOTA was assessed.

Overall conclusion is that all of these models perform similarly. This can likely be attributed to the size of contexts (transcript for each video is fairly small) and their relative diversity, so even the two base models correctly identified the appropriate context and fine-tuning did not bring much benefit.

The Hugging Face live demo runs the fine-tuned model.

8. Future changes

• Expected Changes & Future Improvements:

  1. Re-evaluate Fine-Tuning Strategy: 🤔 Given results, embedding fine-tuning needs review. This could involve:
     • Augmenting the fine-tuning dataset or using different data generation strategies.
     • Changing the semantic chunking strategy to produce more targeted context which may be especially important on edge devices. This could in turn increase the importance of fine tuning.
  2. Prompt Engineering: ✍️ Refine LLM agent prompts (supervisor, RAG) for better answer synthesis. This could boost factual correctness and relevancy, regardless of embedding model.
  3. Advanced RAG Techniques: ✨ Explore methods like re-ranking, query transformations, or HyDE. The goal is to improve context quality and relevance for the LLM.
  4. LLM for Generation: 🧠 Experiment with different LLMs for answer generation. evaluate_rag.ipynb uses gpt-4.1-nano (RAG - for efficiency) and gpt-4.1-mini (evaluator); app.py uses gpt-4.1-mini. Consistency or a more powerful model might improve results.
  5. A more complex agent team. Possibilities:
  • LLM that writes queries for tools based on previous messages,
  • Writing team that can develop a presentation based on the produced research results.
  • A "highlighter" that can identify the object of discussion in the frame and circle it.
• 6. A more complex ingestion pipeline, that is able to transcribe and OCR videos even when they are not accompanied by the transcripts.