Build an End-to-End Modular PDF RAG XML Role Validator

Validating whether roles listed in structured configuration files match job titles described across unstructured project documentation is a common, error-prone manual task. Differences in phrasing, abbreviations, and formatting frequently break simple keyword matching.

This project automates the cross-validation of structured XML specifications against unstructured PDF documents by combining Retrieval-Augmented Generation (RAG) with Fuzzy Logic Algorithms.

System Architecture & Workflow

The pipeline bridges structured parsing, semantic retrieval, and algorithmic string comparison:

[Structured XML File] ──► [lxml Parser] ────────────► [Expected Roles List]
                                                              │
                                                              ▼
[Unstructured PDF Doc] ──► [PyMuPDF + Gemini RAG] ──► [Role Comparer Engine]
                                (Pinecone DB)                 │
                                                              ├─► Levenshtein Distance (Typos)
                                                              └─► Fuzzy Partial Ratio (Abbreviations)
                                                              │
                                                              ▼
                                                    [Streamlit Validation Report]

1. Structured Data Extraction: The application uses lxml to parse structured XML files and extract standardized role definitions and attributes.

2. Unstructured Document Ingestion & RAG: PyMuPDF extracts raw text and tables from PDFs. Sections are chunked, embedded using Google Gemini embeddings, and indexed into Pinecone to semantically retrieve contextual role descriptions.

3. Dual-Layer Role Comparison:

   • Semantic Extraction: Gemini Pro extracts candidate job roles directly from the retrieved PDF context.

   • Algorithmic Fuzzy Matching: The system evaluates candidate roles against the XML baseline using two algorithmic metrics:

     • Levenshtein Distance: Catches direct typographical errors and minor character substitutions (e.g., "Tester" vs. "Teater").

     • Fuzzy Partial Ratio (Ratcliff-Obershelp): Identifies substring matches, acronyms, and common abbreviations (e.g., "Software Engineer" vs. "Software Eng.").

4. Interactive Reporting: A Streamlit frontend displays side-by-side reconciliation reports, flagging matched roles, mismatched titles, and missing definitions.

Core Technology Stack

Key Takeaway

Pure semantic search can sometimes overlook small character errors, while standard string matching fails on paraphrased terms. By combining Gemini-driven RAG retrieval with deterministic fuzzy matching algorithms, the system delivers high accuracy for enterprise document validation workflows.