Nebula

Nebula is an Amazon Nova-powered grant drafting workspace.
It turns source documents into cited draft sections, requirement coverage, and exportable submission artifacts.

Watch the demo on Vimeo: https://vimeo.com/1163998408

What It Does

• Extracts requirements from RFP-like documents
• Generates citation-backed draft sections
• Computes coverage (met / partial / missing)
• Flags missing evidence
• Supports citation click-through evidence inspection and grouped missing-evidence guidance
• Exports JSON + Markdown bundles

Stack

• Frontend: Next.js
• Backend: FastAPI
• Storage (local dev): SQLite + local filesystem
• Storage (prod): RDS Postgres + S3
• Models: Amazon Nova + Titan Embeddings via Bedrock

Deployed AWS Architecture

flowchart TB
  U[User Browser] -->|HTTPS| CF[CloudFront Distribution]
  CF -->|Origin| ALB[Application Load Balancer]

  subgraph ECS["ECS Fargate (nebula-cluster)"]
    FE[nebula-frontend<br/>Next.js]
    BE[nebula-backend<br/>FastAPI]
  end

  ALB -->|/*| FE
  ALB -->|/api/*| BE

  FE -->|fetch /api same origin| CF

  BE -->|SQL over TLS| RDS[(RDS Postgres)]
  BE -->|Objects| S3[S3 Uploads Bucket]
  BE -->|Secrets injection| SM[Secrets Manager<br/>DATABASE_URL]
  BE -->|LLM + Embeddings| BR[Amazon Bedrock<br/>Nova + Titan]

  FE --> CW[CloudWatch Logs]
  BE --> CW

Loading

Quick Start

1. Copy env files:

cp backend/.env.example backend/.env
cp frontend/.env.example frontend/.env.local

2. Run with Docker:

docker compose up --build

3. Open:

• Frontend: http://localhost:3000
• Backend: http://localhost:8000
• Backend docs: http://localhost:8000/docs

AWS Deploy Checklist

• Terraform baseline stack: infra/terraform/aws
• Workflow: .github/workflows/deploy-aws.yml
• Prerequisites and secrets: docs/wiki/AWS-Production-Deployment-Checklist.md
• Readiness check script: scripts/aws/check_deploy_readiness.sh

Authentication

• Frontend auth uses Cognito Hosted UI with Google and OAuth 2.0 authorization code + PKCE.
• Backend enforces bearer token validation when AUTH_ENABLED=true.
• Keep frontend and backend client IDs aligned:
  • Frontend: NEXT_PUBLIC_COGNITO_CLIENT_ID
  • Backend: COGNITO_APP_CLIENT_ID
• Required frontend secrets when auth is enabled:
  • NEXT_PUBLIC_AUTH_ENABLED=true
  • NEXT_PUBLIC_COGNITO_DOMAIN
  • NEXT_PUBLIC_COGNITO_CLIENT_ID
  • NEXT_PUBLIC_COGNITO_REDIRECT_URI
  • NEXT_PUBLIC_COGNITO_LOGOUT_REDIRECT_URI
  • NEXT_PUBLIC_COGNITO_SCOPE (recommended: openid email)

Deploy Notes

• After frontend deploy, invalidate CloudFront paths:
  • /
  • /_next/static/*
• If login fails with invalid_client_secret, the Cognito app client is secret-based. Create a public app client (--no-generate-secret) and update both frontend and backend client IDs.

Core API

• POST /projects
• POST /projects/{id}/upload
• POST /projects/{id}/reindex
• POST /projects/{id}/extract-requirements
• POST /projects/{id}/generate-section
• POST /projects/{id}/generate-full-draft
• POST /projects/{id}/coverage
• GET /projects/{id}/export

Workspace Quality Workflow

After each run, review these workspace panels before final export:

• Quality Signals: parse quality counts, extraction mode, candidate dedupe metrics, and RFP ambiguity warnings.
• Unresolved Coverage Gaps: requirement-level partial / missing items with coverage notes and evidence refs.
• Missing Evidence: grouped upload guidance for unresolved claims.

Recommended flow:

1. Upload clean, text-searchable source files.
2. Run generate and review quality diagnostics.
3. Address flagged gaps with targeted uploads.
4. Re-run until unresolved gaps are cleared.
5. Export markdown/json bundle.

Notes

• Current parser registry supports .txt, .md, .csv, .json, .yaml, .yml, .xml, .html, plus native .pdf, .docx, and .rtf.
• Embedding modes: EMBEDDING_MODE=hash|bedrock|hybrid with BEDROCK_EMBEDDING_MODEL_ID for Bedrock-backed vectors.
• Deterministic requirements extraction uses ordered passes (explicit_tag -> structured_outline -> inline_indicator -> fallback_question) and stores question-level provenance.
• If embedding settings change after indexing, API responses may include warnings with code: embedding_dim_drift; re-index documents to resolve.