Multilingual campus service assistant
A multilingual information retrieval system that answers campus-service questions from approved university documents in English and Hindi, with citations to the original source.
Project definition
Problem statement
University service information is often spread across circulars, PDF notices, department pages, handbooks, and frequently changing announcements. Students may know the question they need to ask without knowing which office or document contains the answer.
A general chatbot can produce a fluent response without reliable evidence. This project instead limits answers to approved campus material, shows the supporting source, and avoids answering when the available evidence is insufficient.
Project objectives
- Answer routine campus-service questions using only approved institutional documents.
- Accept questions in English and Hindi while preserving the meaning of the original source.
- Attach document citations to answers so students can verify important information.
- Give authorised staff a controlled way to add, replace, archive, and re-index documents.
- Measure retrieval quality, citation correctness, response usefulness, and refusal behaviour.
System design
System modules
Document preprocessing
Extracts text from approved PDF and text documents, divides it into traceable passages, and records source metadata.
Language processing
Detects the query language, normalises the input, and supports English and Hindi retrieval through a replaceable multilingual pipeline.
Semantic retrieval
Creates vector representations, finds relevant passages, ranks results, and retains document and page references.
Answer generation
Generates an answer from retrieved passages, attaches citations, requests clarification, and rejects unsupported questions.
Document control
Allows authorised staff to add, version, archive, re-index, and classify university documents.
Evaluation subsystem
Runs prepared questions, records retrieval and response metrics, and stores human-review results.
Methodology
System workflow
- 01Input approved documents
A staff user uploads a notice, handbook, form, or service document with its language, category, owner, and effective date.
- 02Build the retrieval index
The system extracts and segments the content, creates vector representations, and retains references to the original pages.
- 03Process the question
The system detects and normalises a question submitted in English or Hindi.
- 04Retrieve evidence and generate an answer
Relevant passages are ranked and supplied to the answer model. The result includes the documents and pages used.
- 05Evaluate the result
Feedback, low-confidence responses, and prepared test questions are used to calculate and review system performance.
Demonstration scenario
A student asks in Hindi about an active scholarship notice. The assistant returns a concise answer with the relevant document and page citation. The student opens the source, asks a follow-up question, and submits feedback. An administrator then replaces the notice with a newer version and verifies that the assistant uses the updated source.
Engineering
Technical architecture
- Web application
- Nuxt and Vue for the student experience and administration console, with accessible responsive layouts.
- Application API
- FastAPI for document workflows, query handling, access control, feedback, and evaluation endpoints.
- Data layer
- PostgreSQL for users, documents, metadata, feedback, and audit records, with pgvector for semantic retrieval.
- Language pipeline
- Python-based document processing, multilingual embeddings, language detection, and a replaceable translation adapter where required.
- Answer strategy
- Retrieval-augmented generation with citation requirements, confidence thresholds, prompt constraints, and unsupported-answer refusal.
- Delivery
- Container-based local setup and deployment documentation with separate configuration for application secrets and model providers.
Testing
Evaluation
Evaluation measures
- Relevant-passage retrieval rate across the prepared question set
- Citation correctness and source-page traceability
- Groundedness and unsupported-answer refusal rate
- English and Hindi meaning consistency on paired questions
- Response usefulness scored with a documented human-review rubric
- Response latency under the defined demonstration workload
System boundaries
- The assistant does not make admissions, fee, scholarship, disciplinary, or academic decisions.
- It does not search the unrestricted public web or treat unapproved material as university policy.
- It does not expose private student records or require personal information for routine questions.
- Low-confidence or unsupported questions are redirected to the responsible campus office.
Included
- 01Student query interface and document management module
- 02Approved-document ingestion, indexing, retrieval, and citation pipeline
- 03English and Hindi question support with an adapter for additional languages
- 04Prepared evaluation dataset, test results, and demonstration scenarios
- 05Complete source code in a private GitHub repository
- 06Complete project documentation with synopsis, abstract, methodology, architecture diagrams, test results, screenshots, and conclusion
Project record
No buyer information is collected on this page.
- Permanent project ID
- GP-CS-117QU6G
- Catalogued
- 21 Aug 2026
- Completed
- Pending
- Verified
- Pending
- Demonstration
- Added when ready
Handover
After purchase
- 01Payment is confirmed
The project is marked unavailable and cannot be purchased again.
- 02Repository access is granted
The buyer's submitted GitHub account receives access to the private repository.
- 03The purchase record is delivered
The certification sheet is prepared from the reviewed buyer details and sent privately by email.