Parse complex document layouts while preserving page numbers and section headers.
Project-based Internship Programme
Build a Cited Document Assistant in a Generative AI Internship Project
This generative AI online internship with certificate is a fee-based, project-based internship programme. This generative AI online internship with certificate is a fee-based, project-based internship programme that focuses on retrieval-augmented generation (RAG) architectures. You build an intelligent document assistant that ingests complex PDF documents, retrieves semantic evidence, enforces verifiable source citations, and demonstrates honest abstention when evidence is absent.
Decision note 01
Who this project fits and who it does not
A useful fit if…
- You want to build real-world GenAI applications that solve enterprise hallucination problems through retrieval grounding.
- You understand that responsible AI engineering requires traceable citations and measurable evaluation benchmarks.
- You want to master document ingestion, chunking strategies, dense vector embeddings, and similarity search pipelines.
Choose another route if…
- You want to train foundational large language models from scratch on massive supercomputing clusters.
- You only want to generate AI art images; this project focuses on enterprise document retrieval and knowledge extraction.
- You expect language models to guess answers when source documents do not contain the necessary facts.
Official assigned project
AI Document Assistant / RAG
Large language models possess impressive conversational fluency, but in enterprise contexts, ungrounded models frequently invent plausible-sounding falsehoods. In this Generative AI project, you build an AI Document Assistant using Retrieval-Augmented Generation (RAG). You constrain the model's responses to a defined corpus of authorized PDF documents. Every factual statement generated by your assistant must cite the exact page and source passage, and when an answer cannot be proven by the retrieved text, the system must clearly declare that the information is unavailable.
Task brief
AI Document Assistant / RAG - Build a retrieval-augmented assistant with document ingestion, retrieval, grounded responses, and evaluation.
Catalogue deliverables
- Document ingestion pipeline preserving PDF page provenance and document metadata
- Chunking and embedding module with documented token/character boundary configurations
- Vector similarity retrieval service returning ranked context passages with relevance scores
- Grounded conversational interface presenting verifiable citations and honest abstention states
- Evaluation test suite containing supported, unsupported, and ambiguous benchmark questions
How the project works
Grounding language models in verifiable evidence and page-level citations
The engineering challenge begins with document ingestion. Extracting text from PDFs is fraught with formatting anomalies, headers, footers, tables, and multi-column layouts. Your ingestion pipeline parses raw PDF documents while maintaining strict metadata tags for document filename, section titles, and physical page numbers. Preserving this provenance at the extraction stage is essential for generating accurate citations later.
Chunking and embedding bridge unstructured text and semantic search. If text chunks are too small, they lose vital context; if they are too large, retrieval precision degrades and injects irrelevant noise into the language model. You evaluate chunking strategies (such as 500-token chunks with 50-token overlaps) and transform these passages into dense numerical vectors using an embedding model stored within a local vector database.
Retrieval logic must balance recall and precision. When a user submits an inquiry, your service embeds the query and retrieves the top-k most similar passages using cosine similarity. You evaluate whether the retrieved context contains the actual answer, inspect score thresholds, and configure the prompt orchestrator to assemble the evidence into a structured system prompt.
Grounding and abstention form the safety boundary of your assistant. Your prompt engineering instructs the model to answer exclusively from the provided context and format page-level citations. Crucially, you implement and test an explicit abstention path: when the retrieved passages contain insufficient evidence, the assistant gracefully explains that the document corpus does not contain the answer, eliminating hallucinations entirely.
Submission evidence
What makes this work reviewable
- PDF parsing and ingestion pipeline script demonstrating metadata preservation across document pages.
- Vector store setup and configuration documentation detailing embedding dimensions and chunk sizes.
- Retrieval query engine source code showing top-k ranking and similarity score filtering.
- Grounded chat interface source code showcasing inline page citations and honest fallback messages.
- RAG evaluation dataset and benchmark results documenting retrieval precision and answer correctness.
Your build path
Move from question to reviewable evidence
Ingest authorized PDF documents, extracting usable text while preserving page number metadata. Extract clean document text from PDFs while permanently attaching document and page metadata.
Partition text into overlapping chunks and generate dense vector embeddings using a documented model. Chunk passages with structured overlap and generate high-dimensional semantic vector embeddings.
Implement similarity search to retrieve the top matching evidence passages for any user query. Execute similarity searches against your vector index to extract the most relevant context passages.
Construct grounded system prompts requiring page-level citations and enforcing an explicit refusal when context is missing. Direct the language model to synthesize answers grounded strictly in retrieved passages with exact citations.
Evaluate retrieval accuracy and answer faithfulness against a documented benchmark question set. Run automated benchmark evaluations testing retrieval relevance, citation accuracy, and abstention on unanswerable prompts.
Private self-check
Is this project a reasonable learning fit?
Your answers remain in this browser tab and are not stored or sent.
Use these prompts for reflection; they are not an eligibility test.
Skills notebook
Build capability in a realistic order
These are general domain-learning suggestions, not confirmed HireeBridge tool requirements.
Document Processing & Embeddings
Determine optimal token window sizes and overlaps to maintain passage coherence.
Generate dense semantic vectors using OpenAI, Cohere, or HuggingFace embedding models.
Vector Search & Orchestration
Store and query embedded vectors using ChromaDB, FAISS, or Qdrant.
Filter and rank relevant context passages using cosine similarity metrics.
Construct modular RAG pipelines with LangChain or LlamaIndex.
Prompt Engineering & Evaluation
Design system prompts that strictly constrain answers to provided evidence.
Instruct models to return verifiable page references for every stated claim.
Measure context relevance, groundedness, and answer faithfulness against test suites.
Review before submitting
Common Generative AI project mistakes
- 01
Losing page number metadata during PDF parsing
Always store page numbers and source filenames as metadata on each chunk so citations can be traced back to original pages.
- 02
Allowing the model to answer when retrieval finds no relevant text
Configure strict guardrail prompts that instruct the model to state clearly that the answer is not present in the documents.
- 03
Using arbitrary chunk sizes without testing overlap
Excessive chunking fragments ideas across boundaries; test overlap parameters to ensure sentences are not cut mid-thought.
- 04
Evaluating only easy questions where the answer is obvious
Include trick questions, questions outside the document scope, and adversarial prompts to test system robustness.
- 05
Conflating fluency with factual accuracy
A beautifully written answer can still be completely hallucinated; always verify claims against source chunk text.
What reviewers check
Completeness against the assigned brief and deliverables; functional correctness; domain-relevant logic, data, metrics or implementation; required edge cases and failure handling; reproducible setup and submission evidence; and clear documentation of the completed work.
Reviewer
GreyRocks team
Catalogue validation notes
Check page citations, empty or poor retrieval, unsupported questions, and ingestion failures.
Evidence language
Draft an honest CV bullet
Keep placeholders until you can replace them with evidence from your own project.
- Engineered an enterprise RAG document assistant in Python using LangChain and ChromaDB, achieving [metric]% citation accuracy.
- Implemented a PDF ingestion pipeline indexing [number] pages with preserved metadata and zero-hallucination guardrails.
- Optimized semantic vector search across [number] document chunks, improving context retrieval precision by [percentage].
- Constructed a benchmark evaluation suite of [number] queries measuring retrieval relevance and answer faithfulness.
- Built an interactive chat interface in Streamlit displaying real-time source citations and explicit abstention states.
Project readiness
Prepare a strong project submission
Certificate and verification
Completion comes before the credential
GreyRocks provides the technical evaluation and certification authority for HireeBridge programmes. Enrolment and fee payment provide access to the assigned project specifications, sample document sets, and architectural guidelines; they do not automatically issue a certificate. To receive certification, you submit your completed codebase, evaluation benchmark results, and video or screen recording demonstrations. An AI evaluator audits your retrieval grounding, citation integrity, and abstention handling. Approved projects are awarded an official credential with an unalterable ID and QR verification link on GreyRocks.
- Complete
- Submit
- Review
- Approval
- Credential ID and QR
Read the certificate process · Verify a credential on GreyRocks
Duration: 1 Month / 4 Weeks.
Plan inclusions: Each domain maps to an assigned project and task specification. Reference repositories and comprehensive materials depend on the selected plan; certificates follow task submission and explicit reviewer approval.
Questions from students
Generative AI internship FAQ
What specific project will I build in this Generative AI internship?
You will build an AI Document Assistant / RAG project. You will ingest complex PDF documents, create semantic embeddings, store them in a vector database, retrieve relevant context passages for user questions, and generate answers with verifiable page citations while enforcing zero-hallucination abstention.
What is the difference between Generative AI and general Artificial Intelligence?
General AI encompasses classical machine learning, speech recognition, and search algorithms. Generative AI specifically focuses on generative foundation models, prompt engineering, semantic embeddings, and retrieval-augmented architectures that synthesize new text grounded in verified facts.
Do I need to pay for expensive API keys to complete this project?
No. You can complete the project using open-source embedding models (such as HuggingFace sentence-transformers) and local LLMs (via Ollama or Llama.cpp), or by utilizing free-tier developer credits from leading API providers.
What is an abstention path and why is it required?
An abstention path is the system's ability to recognize when the retrieved documents do not contain the answer and explicitly respond that the information is unavailable. This prevents the model from hallucinating false answers.
How do I prove that the citations are real and not fabricated by the model?
Your pipeline extracts and displays the actual text chunks and page numbers returned by the vector database alongside the generated response, allowing an evaluator to compare the generated claim against the source chunk.
Can I use my own choice of PDF documents?
Yes. You can use authorized public sample PDFs (such as technical manuals, research papers, or open textbooks) and document your test queries against that specific collection.
How long does it take to complete the programme?
The curriculum is designed for 4 weeks of structured, self-paced development, covering ingestion, embedding, vector search, interface construction, and evaluation.
Is the certificate recognized for college placement and resume building?
Yes. The certificate is issued upon explicit project review by GreyRocks, providing a verifiable credential ID and QR verification link that demonstrates authentic, hands-on engineering capability.
Next step
Choose your plan and start building.
Review plan details, included resources and the assigned project scope before you begin.