OFC — Offline Report Composer

Air-gapped report writing with local LLMs, file uploads, and database context/examples.

Why

In locked-down networks you often have:

• a private OpenAI-compatible endpoint (Ollama, vLLM, LM Studio, internal gateway)
• result data in a DB or files
• prior reports that should set tone/structure

OFC wires those together into a small pipeline + UI. No cloud fallbacks, no CDN at runtime.

Architecture

UI (React)  →  API (FastAPI)  →  LLM endpoint (your LAN)
                    ↓
              SQLite app store
                    ↓
         uploads / external DBs (read-only SELECT)

Pipeline stages: style notes → outline → draft → critique → revise (or run one stage at a time). When examples are attached, style notes capture formulation signals (phrases, voice, section naming, metrics/hedging) so later stages prefer that language without copying example-only facts. Style notes are cached on disk per example-set fingerprint (data/style_cache/) so identical packs skip re-extraction. Changing library examples, example uploads, or example SQL queries clears that cache and stored notes on reports; changing a report’s example selection or brief (when “use all examples” ranks the pack) clears that report’s notes when the fingerprint shifts.

Training vs retrieval (small corpora)

Continuously fine-tuning / “training” on every stored report is a poor default under ~1000 documents: costly, brittle offline, and needs an eval harness. Prefer attaching the best example reports, the style-notes stage, and the cheap style-notes cache. When “use all examples” is on, OFC ranks library examples by cheap lexical similarity to the title/brief (and keeps a top budget) — not model-weight training.

Quick start

# 0) Config
cp .env.example .env   # already present if you cloned this tree

# 1) Sample archive DB (results + prior reports)
python3 scripts/seed_sample_db.py

# 2) API
cd backend
python3 -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt
OFC_DATA_DIR=../data uvicorn app.main:app --reload --host 0.0.0.0 --port 8000

# 3) UI (another terminal)
cd frontend
npm install
npm run dev

Open http://localhost:5173

Point LLM_* in .env at your local model server first (default assumes Ollama on 11434).

Docker

docker compose up --build

• UI: http://localhost:8080
• API: http://localhost:8000

For true air-gap: build on a connected host, docker save the images, transfer, docker load.

Typical workflow

1. LLM & settings — add a profile (base URL, model, API key) or rely on .env
2. Sources
   • Upload .md / .txt / .csv / .json / .docx as context, example, or both
   • Add a DB connection (SQLite path or Postgres/MySQL DSN) and saved SELECT queries
     • purpose=results → metrics/tables for the new report
     • purpose=examples → prior report bodies (example_body_column)
   • Review the document library: short smoke/placeholder entries are flagged as stubs — prune them so they do not dilute example packs
3. Reports — create a project, write the brief, tick files/queries, run Full generate (runs in the background with live stage status; Cancel aborts the in-flight LLM HTTP call). The editor warns early if the configured LLM is unreachable or lists no models. If a draft still echoes example-only phrasing, use Scrub example phrasing
4. Optionally Import theme from a .docx example or library doc (fonts, header/footer, logos); live preview mirrors that chrome, and Word/HTML export apply it on download
5. Edit the draft; export Markdown, HTML, or .docx

Sample DB hints

After scripts/seed_sample_db.py:

Field	Value
dialect	sqlite
dsn	absolute path to data/samples/archive.sqlite3
results SQL	SELECT metric, value, unit, period, notes FROM quarterly_results
examples SQL	SELECT title, year, body_md FROM prior_reports
example body column	body_md

Air-gap notes

• Frontend production build is self-contained (no runtime CDN)
• App state + uploads live under OFC_DATA_DIR (default ./data)
• External DB credentials are Fernet-encrypted using OFC_SECRET_KEY / OFC_FERNET_KEY
• SQL is restricted to read-only SELECT / WITH
• LLM traffic stays on the URLs you configure — nothing phones home

Config reference

Variable	Purpose
LLM_BASE_URL	OpenAI-compatible root, e.g. http://127.0.0.1:11434/v1
LLM_API_KEY	Bearer token (placeholder OK for many local servers)
LLM_MODEL	Model id
OFC_DATA_DIR	Uploads, SQLite app DB, exported reports
OFC_SECRET_KEY	Derives encryption key for stored secrets
OFC_CORS_ORIGINS	Allowed browser origins

Layout

ofc/
  backend/app/          FastAPI + pipeline
  frontend/             Vite React UI
  data/                 Local store (gitignored contents)
  data/samples/         Demo SQLite + hint file
  docker/               API + nginx UI images
  scripts/              Seed helpers