Skip to content

Latest commit

 

History

History
257 lines (173 loc) · 8.14 KB

README.md

File metadata and controls

257 lines (173 loc) · 8.14 KB

Semantic-Embedding-for-Mood-Based-Yoga-Session-Recommendation-and-Report-Generation-Proof-of-Concept

This repo shows the codes for Semantic Embedding for Mood-Based Yoga Session Recommendation and Report Generation

Ollama Installation

curl -fsSL https://ollama.com/install.sh | sh

Then use required LLMs and Embeddings from Ollama

Setup

  1. Python3 venv -m yoga
  2. source yoga/bin/activate
  3. pip install -r requirements

Run the Server Code on FastAP

Python3 mood_9.py

Run the Client Code

From same project directory run

python3 curl_caller_mood.py

Result

  1. A directory called reports is generated with PDF reports based on mood which was given in the prompts as mentioned in curl_caller_mood.py
  2. A CSV file for metrics analysis is generated yoga_asana_metrics.csv

Detailed Step-by-Step Workflow

1. Application Setup

1.1 Initialize FastAPI Application

  • Instantiate a FastAPI object to manage routing and endpoint handling.
  • Define endpoints for processing user prompts and serving PDF reports.

1.2 Configure Logging

  • Set logging to the DEBUG level to capture detailed logs.
  • Use StreamHandler to ensure logs are visible in the console during runtime.
  • Log all key operations and errors for debugging and tracking.

1.3 Define Configuration Variables

  • Embedding Model: Choose a pre-trained embedding model (mxbai-embed-large) to compute vector representations of text.
  • Similarity Threshold: Set a similarity threshold (0.62) to determine if an asana matches the user prompt.
  • API Endpoints:
    • Ollama API for embedding generation.
    • Ollama API for generating comments via a large language model (LLM).
  • Report Directory: Create a folder named reports to store generated PDF reports.
  • CSV Metrics File: Define a CSV file (yoga_asana_metrics.csv) to log request metrics.

2. Data Initialization

2.1 Predefined Yoga Asanas

  • Define a list of yoga asanas, each containing:
    • Name: Name of the asana.
    • Utterances: Sample phrases representing moods the asana addresses.
    • Details:
      • Step-by-step instructions.
      • Recommended frequency and timing.
      • Dietary and lifestyle recommendations.
      • Benefits.

2.2 Embedding Cache

  • Create a dictionary (cached_embeddings) to store embeddings of predefined asana utterances.
  • Use the cache to avoid redundant API calls for embeddings.

3. Startup Tasks

3.1 Load Embedding Model

  • On application startup:
    • Test the embedding API by generating an embedding for a sample text.
    • Log success or failure of the embedding model.

3.2 Cache Asana Utterance Embeddings

  • Precompute embeddings for all utterances in the predefined asanas using the embedding API.
  • Store these embeddings in the cached_embeddings dictionary for efficient retrieval during requests.

4. Request Handling

4.1 Endpoint: /process_prompt

Step 1: Receive User Input

  • Accept a JSON payload containing a prompt (e.g., "I feel overwhelmed and stressed with my work").
  • Parse and validate the input using Pydantic.

Step 2: Generate Prompt Embedding

  • Send the user prompt to the embedding API to retrieve its semantic vector representation.
  • If the embedding generation fails, return a 500 Internal Server Error response.

Step 3: Find the Best Asana

  • Iterate through the predefined asanas and compute the cosine similarity between the prompt embedding and the cached embeddings of asana utterances.
  • Select the asana with the highest similarity score above the threshold (0.62).
  • Measure and log the time taken for this matching process.
  • No Match Handling:
    • If no asana meets the similarity threshold:
      • Log the event with no_route.
      • Save metrics to the CSV file.
      • Return a "no match" response to the user.

Step 4: Generate Final Comment

  • Use the LLM API to generate a personalized comment about the selected asana.
  • Pass a few-shot example (predefined input-output pairs) and the selected asana's details to the API for context.
  • If the LLM API fails, use a fallback message and log the error.

Step 5: Generate PDF Report

  • Use the ReportLab library to create a PDF report containing:
    • User prompt and similarity score.
    • Detailed instructions for the recommended asana.
    • Dietary and lifestyle recommendations.
    • LLM-generated final comment.
  • Save the PDF in the reports directory.
  • If PDF generation fails, return a 500 Internal Server Error response.

Step 6: Log Metrics

  • Record the following details in the CSV file:
    • Date and time of the request.
    • Selected asana and similarity score.
    • LLM metrics (response time, tokens processed, etc.).
    • PDF generation time and embedding match duration.

Step 7: Return Response

  • Respond to the user with:
    • Selected asana details.
    • LLM-generated final comment.
    • Download link for the PDF report.
    • Total processing time.

5. Report Download

5.1 Endpoint: /download_report/{report_filename}

  • Accept a filename as input.
  • Locate the corresponding PDF file in the reports directory.
  • Return the file as a downloadable response.
  • Error Handling:
    • Return a 404 Not Found response if the file doesn’t exist.

6. Utility Functions

6.1 Embedding Generation

  • Send a POST request to the embedding API with the prompt or utterance.
  • Parse and return the embedding vector from the API response.

6.2 Cosine Similarity Calculation

  • Compute the similarity between two embeddings using: [ \text{Cosine Similarity} = \frac{\text{dot product of vectors}}{\text{product of vector magnitudes}} ]

6.3 Find Best Asana

  • Compare the prompt embedding against cached embeddings for all asana utterances.
  • Return the asana with the highest similarity score above the threshold.

6.4 Generate Final Comment

  • Send the selected asana's details and user prompt to the LLM API.
  • Use few-shot learning to guide the LLM in generating concise, context-aware comments.

6.5 Generate PDF Report

  • Use the ReportLab library to create a visually structured PDF with:
    • Header: Mood-Based Yoga Recommendation Report.
    • User prompt and similarity score.
    • Detailed instructions and recommendations for the asana.
    • Final comment.
  • Save the report in the reports directory.

7. Metrics Logging

  • Record key metrics for each request:
    • Request Details: User prompt, selected asana.
    • Performance Metrics:
      • Embedding match duration (logged as embed_match_duration in nanoseconds).
      • LLM response time and tokens processed.
      • PDF generation time.
    • Save these metrics to the yoga_asana_metrics.csv file.

8. Error Handling

  • Handle exceptions at every step:
    • Log errors with detailed messages.
    • Return user-friendly HTTP error responses:
      • 500 for internal server issues.
      • 404 for missing files.

9. Application Deployment

  • Start the FastAPI server using uvicorn on port 5000.
  • Test endpoints using tools like curl or Postman to ensure end-to-end functionality.

Sample PDF Report on Mood Based Yoga Asana

image

For Web Integration

Look into the web_integration directory and run the codes accordingly. The web integration uses gradio and sqlalchemiy to use a local database “yoga_interactions.db” having columns such as: id, timestamp, prompt, final_response, pdf_file_path

When ready

image

When got response

image

Sometimes, if output is not suitable, just re-SUBMIT

Example

Example query to be given as input into the gradio:

Modify as per Detailed_Mood_Issue column of mood_asana_details.csv file

• I'm feeling guilty and remorseful.

• Irritability is affecting my mood.

• Dealing with restlessness and impatience.

• Agitation is making me restless.

• I'm feeling lonely and isolated.

etc…