https://github.com/sourish-kanna/sdp-llm

https://github.com/sourish-kanna/sdp-llm

Last synced: 3 months ago
JSON representation

• Host: GitHub
• URL: https://github.com/sourish-kanna/sdp-llm
• Owner: Sourish-Kanna
• Created: 2025-07-08T16:42:32.000Z (3 months ago)
• Default Branch: main
• Last Pushed: 2025-07-12T15:05:21.000Z (3 months ago)
• Last Synced: 2025-07-12T17:31:04.177Z (3 months ago)
• Size: 2.93 KB
• Stars: 0
• Watchers: 0
• Forks: 0
• Open Issues: 0
• Metadata Files:
  • Readme: README.md

Awesome Lists containing this project

README

          
# 📟 Autonomous Audit+ – A Hybrid Agentic Financial Document Auditor

## 🔍 Overview

Autonomous Audit+ is a hybrid multi-agent invoice auditing system that combines rule-based logic with natural language understanding. It processes both CSV and PDF invoices to:

* Detect financial issues

* Generate plain-English explanations

The system uses:

* **Mistral 7B via Groq + LangChain** for rule-based and RAG audit logic

* **Gemini 1.5 Pro** for fluent summaries and suggestions

* **Streamlit** as the UI frontend

---

## 👥 Team Role Breakdown

### 👥 Frontend Team

#### 👤 Parth – UI/UX + File Input

* Design and implement Streamlit UI layout

* Build file upload components for CSV and PDF

* Display parsed invoices using Pandas tables

* Style and improve visual clarity and responsiveness

#### 👤 Parth & Sourish – Integration + Output Display

* Connect Mistral and Gemini backends to frontend

* Trigger audit and explanation functions from UI

* Render output summaries and issues

* Integrate charts and export options (PDF/Markdown)

### 👤 Sourish – Gemini + Input / Embeddings

* CSV and PDF file parsing

* Normalizing data into a unified DataFrame

* Chunking and embedding documents using LangChain-compatible vector store

* Generating Gemini prompts for:

  * Audit summary

  * Multi-agent role simulation (accountant, legal, manager)

  * Suggestions

* Exporting markdown/PDF reports

### 👤 Prem – Mistral + Audit Logic

* Build LangChain tools for audit rules:

  * Total mismatch

  * Missing fields

  * Duplicate invoice IDs

  * Future dates

* Set up Mistral via Groq in LangChain agent

* Integrate with retriever provided by you

* Implement RAG (Retrieval-Augmented Generation) chains

* Output structured JSON for Gemini processing

---

## 📁 Folder Structure

```text

invoice_audit_agent/

├── app.py                    # Streamlit app UI

├── embeddings/               # Embedding logic 

│   ├── vector_store.py

│   └── retriever_utils.py

├── parsers/                  # PDF/CSV handling 

│   ├── csv_parser.py

│   └── pdf_parser.py

├── reports/                  # Gemini prompt + writer 

│   └── gemini_writer.py

├── audit/                    # Mistral logic 

│   ├── audit_tools.py

│   ├── audit_agent.py

│   └── rag_agent.py

├── utils/

│   └── formatter.py          # Shared helpers

├── sample_data/              # Test files

├── requirements.txt          # Dependencies

└── .env                      # Secrets/API keys

```

---

## 📄 File Format Guidelines

### ✅ Initial Normalized DataFrame (After Parsing CSV/PDF)

| Field       | Type                |

| ----------- | ------------------- |

| invoice\_id | string              |

| vendor      | string              |

| date        | string (YYYY-MM-DD) |

| quantity    | float               |

| unit\_price | float               |

| total       | float               |

This is parsed from CSV or PDF, and passed to:

* Vector embedder (Sourish)

* Mistral audit agent (Prem)

---

### 🔄 Intermediate Format to Mistral (Shared Input)

```json

{

  "invoices": [

    {

      "invoice_id": "INV-101",

      "vendor": "ABC Traders",

      "date": "2025-01-23",

      "quantity": 250,

      "unit_price": 110.0,

      "total": 27500.0

    },

    ...

  ]

}

```

* This structure is passed from **Sourish to Prem**.

* Also stored in vector DB.

---

### 📉 Output Format from Mistral to Sourish (for Gemini)

```json

{

  "invoice_summary": {

    "total_invoices": 12,

    "files_processed": ["jan.csv", "feb.pdf"]

  },

  "issues": [

    {

      "invoice_id": "INV-1024",

      "vendor": "ABC Traders",

      "issue_type": "Total Mismatch",

      "description": "Total does not match quantity x unit price",

      "severity": "high"

    }

  ],

  "compliance_flags": {

    "gst_issues": 3,

    "future_dates": 1

  },

  "top_vendors": [

    { "name": "ABC Traders", "amount_billed": 18500 }

  ]

}

```

* Returned from **Prem to Sourish**

* Used for Gemini summaries and markdown report generation

---

## 🗓️ Development Timeline

| Day | Task                                   |

| --- | -------------------------------------- |

| 1   | UI + file upload                       |

| 2   | File parsing + cleaning                |

| 3   | Audit rule logic (mistral)             |

| 4   | LangChain agent setup                  |

| 5   | Gemini summary generation              |

| 6   | Role-based agent views                 |

| 7   | Duplicate detection (cross-invoice)    |

| 8   | Charting and final audit summary       |

| 9   | Internal testing                       |

| 10  | Polish, test, and export final reports |

---

## 🚀 Next Steps

* [ ] Frontend Member 1: Finalize UI layout and file upload modules

* [ ] Frontend Member 2: Integrate Mistral/Gemini pipelines and display results

* [ ] Backend (Prem): Build and test audit tools, connect RAG model

* [ ] Gemini + Embedding (Sourish): Finalize file parsing, embed data, create Gemini prompt chains

* [ ] All: Test integration flow from input → audit → summary → export

---

> This README serves as the central reference document for planning, development, and integration. Keep it updated as modules evolve.