File upload support: PDF, docx, markdown for admin and participant inputs

[Par-t]

Problem

Currently only plain text input is supported for both admin (hackathon setup) and participant (idea submission). Real-world usage requires document uploads — hackathon rules as PDFs, ideas written in docx/markdown, etc.

Current state

• Admin: text chat with DeepSeek via /init endpoint
• Participant: idea_text: str field in submission JSON
• No file upload endpoint exists
• Frontend will have a file input prop ready (non-functional until backend support lands)

Proposed solution

Participant side (ingestion/normalization node)

• Accept file uploads (PDF, docx, markdown) as base64 in submission payload
• Ingestion node extracts text based on file type → conditionally summarizes if too long
• Normalized text feeds into embeddings + scoring
• Tools: extract_pdf, extract_docx, parse_markdown — agent picks the right one based on input

Admin side (init handler)

• Accept file upload (hackathon rules/guidelines document) in /init payload
• Quick text extraction → keyword scan for criteria/guidelines/theme sections
• Send relevant sections to LLM for structured config extraction
• If document is missing required info, ask admin for clarification in follow-up turn

Libraries

• PDF: pdfplumber
• Docx: python-docx
• Markdown: regex strip or mistune

Scope

• Add idea_file + idea_file_type optional fields to submission model
• Add file + file_type optional fields to init request
• Extraction tools as agent-callable functions inside TEE
• Conditional summarization for long extracted text
• Error handling for malformed/empty files

Context

Part of making the pipeline genuinely agentic — the ingestion node makes non-deterministic tool call decisions based on input format and content length. Also unblocks real-world usage where participants have their ideas in documents rather than raw text.