Content Factory

Voice-to-content pipeline with six content types, two voice registers, A/B testing, AI scoring, and SQLite persistence. Record a voice note or type text, get a structured draft (title, tags, formatted body), revise with another voice or text note, compare voice register variants in a sequential A/B review flow, then save as markdown with YAML frontmatter.

Source ownership note:

Application source lives in /mnt/apps/src/factory/services/content-factory
This runtime directory owns deployment files, persistent runtime data, and service orchestration

Quick Start

cd /mnt/apps/docker/ai/content-factory
docker compose up -d --build

# View logs
docker compose logs -f

Access: https://factory.haiven.site

Architecture

Browser (voice or text input)
    |
    | audio blob (WebM / MP4 fallback for Safari)
    v
Content Factory API (FastAPI, port 8020)
    |
    |--- POST /classify ---------> haiven-transcribe (STT)
    |                                   |
    |                                   v
    |                               raw transcription + AI content type suggestions
    |                               (classify session stored in-memory, 1-hour TTL)
    |
    |--- POST /classify-text ----> classify_content() ---> LiteLLM ---> GLM-4.7-Flash
    |
    |--- POST /format -----------> formatter.py ---> LiteLLM ---> Hermes-4.3-36B
    |                                   |
    |                                   v
    |                               JSON (title, tags, formatted_text)
    |                               session persisted to SQLite
    |
    |--- POST /revise -----------> voice feedback ---> transcribe + revise
    |--- POST /revise-text ------> text feedback ---> revise
    |
    |--- POST /save -------------> writer.py ---> /content/drafts/{type}/{date}-{slug}.md
    |                               draft index record written to SQLite
    |
    |--- GET /drafts ------------> list saved drafts (filesystem scan)
    |--- GET /drafts/{path} ----> single draft (path traversal protected)
    |--- GET /types -------------> list content type definitions
    |
    |--- POST /ab-tests ---------> db.create_ab_test() ---> SQLite
    |--- GET  /ab-tests ---------> db.list_ab_tests() (supports ?group_by=voice_pair)
    |--- GET  /ab-tests/{id} ----> test detail + scores + metadata + highlights
    |--- POST /ab-tests/{id}/autosave    ---> partial progress (every 120s)
    |--- PATCH /ab-tests/{id}            ---> update name/status/metadata
    |--- GET  /ab-tests/{id}/scores      ---> human + AI scores
    |--- DELETE /ab-tests/{id}           ---> soft-delete (archive)
    |
    |--- POST /ab-tests/{id}/ai-score    ---> ai_score.py ---> LiteLLM ---> GLM-4.7-Flash
    |--- POST /ab-tests/generate-name    ---> ai_score.py ---> GLM-4.7-Flash
    |
    |--- POST /ab-tests/{id}/highlights  ---> save inline phrase highlights
    |--- GET  /ab-tests/{id}/highlights  ---> retrieve highlights
    |
    |--- POST /ab-scores --------> legacy flat-file score submission (still works)

Components

Module	Purpose
`app/main.py`	FastAPI application, all route definitions
`app/db.py`	Async SQLite layer — 6 tables, WAL mode, CRUD for sessions/drafts/tests/scores/highlights/metadata
`app/migrate.py`	One-time JSON→SQLite data migration (runs on startup if needed)
`app/ai_score.py`	AI scoring and test name generation via GLM-4.7-Flash (thinking disabled)
`app/transcriber.py`	haiven-transcribe integration (OpenAI-compatible STT)
`app/formatter.py`	Content-type-specific prompt building, Seed-36B via LiteLLM, JSON response parsing, content classification
`app/writer.py`	Markdown file writer with YAML frontmatter, slug generation
`app/prompts/content_types.py`	Six content type definitions with voice profiles and anti-pattern lists
`app/static/`	Vanilla HTML/CSS/JS frontend (Crimson Edge theme, hash routing, mobile-first)

Content Types

Type	Purpose
Entry	Personal journal entry or reflection
TIL	Today I Learned — concise technical or factual discovery
Link	Link post with commentary
Quote	Notable quote with context
Build	Build log, project update, or in-progress note
Site Page	Static site page content

Voice Registers

Register	Style
Authority	Direct, declarative, no hedging — for instructional and factual content
Conversational	Natural first-person voice — for entries, reflections, and personal notes

Agent Identity

The service ships with SYSTEM_PROMPT.md — the agent's identity file that defines how the LLM interprets voice input and formats content. This is loaded by formatter.py at startup and sets the tone, priorities, and anti-patterns for all six content types.

SQLite Schema

The database (factory.db) has 6 tables:

Table	Purpose
`sessions`	Active draft sessions (replaces in-memory dict) — id, content_type, data JSON, raw_transcription, revision_count
`drafts`	Index of saved draft files — path, title, content_type, tags, status
`ab_tests`	A/B test records — id, name, input_text, candidate paths, skill_version, model, status
`scores`	Human and AI scores per test+candidate+dimension — score (1-5), comment, source (human/ai)
`highlights`	Inline phrase annotations — start/end offsets, sentiment (positive/negative), text_preview
`test_metadata`	Key-value store for arbitrary test metadata (voice pairs, autosave state, etc.)

Service Configuration

Property	Value
Container Name	`content-factory`
Image	Custom build (`python:3.12-slim`)
Internal Port	8020
External Port	8023
Domain	`factory.haiven.site`
Networks	`web`, `backend`
GPU	None (CPU only)
Restart Policy	unless-stopped

Docker Compose

services:
  content-factory:
    build:
      context: /mnt/apps/src/factory/services/content-factory
      dockerfile: /mnt/apps/docker/ai/content-factory/Dockerfile
    container_name: content-factory
    restart: unless-stopped
    ports:
      - "8023:8020"
    networks:
      - web
      - backend
    volumes:
      - ${CONTENT_PATH:-/mnt/apps/src/factory/content}:/content
      - /mnt/apps/src/factory/services/content-factory/app:/app/app
      - ./data:/data
      - /mnt/apps/src/factory/.claude/skills/elijah-voice-skill/references:/voice-profiles:ro
      - /mnt/apps/src/factory/docs:/source-docs:ro
    environment:
      TZ: America/New_York
      WHISPER_URL: http://haiven-transcribe:8000/v1/audio/transcriptions
      LITELLM_URL: http://litellm:4000
      LITELLM_KEY: ${LITELLM_KEY:-sk-haiven}
      FACTORY_MODEL: ${FACTORY_MODEL:-hermes-4.3-36b}
      CLASSIFY_MODEL: ${CLASSIFY_MODEL:-qwen3.5-27b}
      CONTENT_PATH: /content
      LANGFUSE_ENABLED: ${LANGFUSE_ENABLED:-true}
      SCORING_MODEL: ${SCORING_MODEL:-glm-4-7-flash}
    labels:
      - "traefik.enable=true"
      - "traefik.docker.network=web"
      - "traefik.http.routers.content-factory.rule=Host(`factory.haiven.site`)"
      - "traefik.http.routers.content-factory.entrypoints=websecure"
      - "traefik.http.routers.content-factory.tls=true"
      - "traefik.http.services.content-factory.loadbalancer.server.port=8020"
    healthcheck:
      test: ["CMD", "python", "-c", "import urllib.request; urllib.request.urlopen('http://localhost:8020/health')"]
      interval: 30s
      timeout: 10s
      retries: 3
      start_period: 15s

networks:
  web:
    external: true
  backend:
    external: true

Environment Variables

Variable	Default	Description
`TZ`	`America/New_York`	Timezone for timestamps in filenames
`WHISPER_URL`	`http://haiven-transcribe:8000/v1/audio/transcriptions`	STT endpoint (OpenAI-compatible)
`LITELLM_URL`	`http://litellm:4000`	LiteLLM gateway for LLM calls
`LITELLM_KEY`	`sk-haiven`	API key for LiteLLM
`FACTORY_MODEL`	`hermes-4.3-36b`	Model used for content drafting and revision
`CLASSIFY_MODEL`	`qwen3.5-27b`	Model used for content type classification suggestions
`SCORING_MODEL`	`glm-4-7-flash`	Model used for AI scoring and test name generation
`CONTENT_PATH`	`/content`	Container path for draft output directory
`DB_PATH`	`/data/factory.db`	SQLite database file location
`LANGFUSE_ENABLED`	`true`	Attach Langfuse trace metadata to LLM calls

Volumes

Host Path	Container Path	Purpose
`/mnt/apps/src/factory/content`	`/content`	Draft output directory (organized by content type)
`./data`	`/data`	SQLite database and A/B test artifacts
`/mnt/apps/src/factory/services/content-factory/app`	`/app/app`	Live source mount for application code
`/mnt/apps/src/factory/.claude/skills/elijah-voice-skill/references`	`/voice-profiles`	Voice profile references (read-only)
`/mnt/apps/src/factory/docs`	`/source-docs`	Source content for hallucination checking (read-only)

Networks

Network	Subnet	Purpose
`web`	10.10.0.0/24	Traefik reverse proxy access
`backend`	10.10.1.0/24	Internal communication with haiven-transcribe and LiteLLM

API Endpoints

`GET /`

Serve the single-page frontend.

Response: HTML document

`GET /health`

Health check for Docker healthcheck and monitoring.

Response:

{"status": "healthy", "service": "content-factory"}

`GET /types`

List all content type definitions with metadata.

Response:

[
  {
    "id": "entry",
    "label": "Entry",
    "description": "Personal journal entry or reflection",
    "registers": ["authority", "conversational"]
  }
]

`POST /classify`

Accept audio, transcribe via haiven-transcribe, and return AI-suggested content types. Stores the transcription in a classify session (1-hour TTL) for use by a subsequent /format call.

Request:
- Content-Type: multipart/form-data
- Body: file field with audio blob (WebM or MP4)

Response:

{
  "transcription": "today i was working on the embedding pipeline...",
  "suggestions": ["til", "build"],
  "classify_id": "a3f1c2d4-..."
}

Status Codes:
- 200 — Success
- 400 — Invalid file type or empty transcription
- 500 — Transcription service error

`POST /classify-text`

Accept typed or pasted text and return AI-suggested content types. Also returns a classify_id for use by /format.

Request:

{"text": "today i was working on the embedding pipeline..."}

Response: Same structure as /classify

`POST /record`

Accept audio, transcribe, format using a specified content type, persist session to SQLite, and return the structured draft.

Request:
- Content-Type: multipart/form-data
- Fields: file (audio blob), content_type (query param, default "article")

Response:

{
  "_draft_id": "b4e2...",
  "title": "embedding dimensions must match at collection creation time",
  "tags": ["qdrant", "embeddings", "debugging"],
  "formatted_text": "Qdrant collections lock their vector dimensions..."
}

`POST /format`

Format text into a structured draft. Accepts either raw text or a classify_id from a prior /classify call. Persists session to SQLite.

Request:

{
  "text": "today i was working on the embedding pipeline...",
  "content_type": "til"
}

Or using a classify session:

{
  "classify_id": "a3f1c2d4-...",
  "content_type": "til"
}

Response: Same structure as /record response (includes _draft_id)

Status Codes:
- 200 — Success
- 400 — Missing required fields or unsupported content type
- 404 — Classify session not found or expired
- 500 — LLM service error

`POST /revise`

Accept voice feedback on an existing draft, transcribe, and return a revised draft. Increments revision count in SQLite.

Request:
- Content-Type: multipart/form-data
- Fields: file (audio), draft_id (query param)

Response: Same structure as /format

`POST /revise-text`

Accept text feedback on an existing draft and return a revised draft.

Request:

{
  "draft_id": "b4e2...",
  "feedback": "make it shorter and add a code example"
}

Response: Same structure as /format

`POST /save`

Write approved draft to the content directory as markdown with YAML frontmatter. Writes a draft index record to SQLite and cleans up the session.

Request:

{"draft_id": "b4e2..."}

Response:

{
  "file_path": "/content/drafts/til/2026-03-15-embedding-dimensions-must-match.md",
  "status": "draft"
}

Status Codes:
- 200 — Success
- 404 — Draft session not found
- 500 — File write error

`GET /drafts`

List saved drafts from the content directory.

Response:

[
  {
    "filename": "2026-03-15-embedding-dimensions-must-match.md",
    "content_type": "til",
    "date": "2026-03-15",
    "title": "embedding dimensions must match at collection creation time",
    "tags": ["qdrant", "embeddings", "debugging"],
    "preview": "Qdrant collections lock their vector dimensions at creation..."
  }
]

`GET /drafts/{path}`

Get a single draft by path. Path is validated against CONTENT_PATH to prevent traversal.

Response: Full draft content including frontmatter fields and raw markdown body.

Status Codes:
- 200 — Success
- 400 — Path traversal attempt
- 404 — Draft not found

`POST /ab-tests`

Create a new A/B test record.

Request:

{
  "name": "Authority vs Conversational — embedding TIL",
  "input_text": "today i was debugging qdrant...",
  "candidate_a": {
    "path": "til/2026-03-15-embedding-dimensions.md",
    "text": "Qdrant collections lock their vector dimensions...",
    "voice_pair": "authority"
  },
  "candidate_b": {
    "path": "til/2026-03-15-embedding-dimensions-conv.md",
    "text": "So I hit a fun bug today — Qdrant...",
    "voice_pair": "conversational"
  },
  "skill_version": "1.0",
  "model": "seed-oss-36b"
}

Response:

{
  "id": "c9f3...",
  "status": "created",
  "name": "Authority vs Conversational — embedding TIL",
  ...
}

`GET /ab-tests`

List all A/B tests with scores and metadata.

Query Parameters:
- group_by=voice_pair — Group results by voice register pairing

Response (default): Array of test objects with scores and metadata keys.

Response (?group_by=voice_pair):

{
  "grouped": {
    "authority vs conversational": [...]
  }
}

`GET /ab-tests/{id}`

Get A/B test detail including candidates, scores, metadata, and highlights.

Response:

{
  "id": "c9f3...",
  "name": "Authority vs Conversational — embedding TIL",
  "status": "completed",
  "candidate_a_path": "...",
  "candidate_b_path": "...",
  "scores": [...],
  "metadata": {"candidate_a_voice_pair": "authority", ...},
  "highlights": [...]
}

`POST /ab-tests/{id}/autosave`

Save partial scoring progress. Called automatically by the frontend every 120 seconds during review.

Request:

{
  "step": "review_a",
  "scores": {"voice": {"score": 4, "comment": "clean"}, "tone": {"score": 3}},
  "timestamp": "2026-03-16T14:22:00"
}

Response:

{"status": "saved", "test_id": "c9f3...", "step": "review_a"}

`PATCH /ab-tests/{id}`

Update test name, status, candidate paths, skill version, or model.

Request:

{"name": "Updated test name", "status": "completed"}

Response:

{"status": "updated", "test_id": "c9f3..."}

`GET /ab-tests/{id}/scores`

Get all scores for an A/B test, both human and AI.

Response:

{
  "test_id": "c9f3...",
  "scores": [
    {
      "candidate": 1,
      "dimension": "voice",
      "score": 4,
      "comment": "direct, no hedging",
      "source": "human"
    },
    {
      "candidate": 1,
      "dimension": "voice",
      "score": 5,
      "comment": "authority register maintained throughout",
      "source": "ai"
    }
  ]
}

`DELETE /ab-tests/{id}`

Soft-delete an A/B test (sets status to "archived"). Data is retained in SQLite.

Response:

{"status": "archived", "test_id": "c9f3..."}

`POST /ab-tests/{id}/ai-score`

Run AI scoring on both candidates via GLM-4.7-Flash (thinking disabled). Scores are persisted to the scores table with source="ai". Candidate texts are pulled from metadata or loaded from draft files.

Response:

{
  "candidate_a": {
    "voice": 5,
    "voice_rationale": "Declarative throughout. No hedging.",
    "arguments": 4,
    "arguments_rationale": "Clear claim, well-supported.",
    "tone": 4,
    "tone_rationale": "Matches authority register."
  },
  "candidate_b": {
    "voice": 3,
    ...
  }
}

Status Codes:
- 200 — Success
- 400 — Both candidate texts required
- 404 — Test not found
- 500 — AI scoring failed

`POST /ab-tests/generate-name`

Generate AI name suggestions for an A/B test based on source text.

Request:

{"text": "today i was debugging qdrant dimension mismatch..."}

Response:

{"suggestions": ["Qdrant Dimension Debug — TIL", "Embedding Mismatch Discovery", "Silent Upsert Failure"]}

`POST /ab-tests/{id}/highlights`

Save an inline phrase highlight for a test candidate.

Request:

{
  "candidate": 1,
  "start_offset": 42,
  "end_offset": 67,
  "sentiment": "positive",
  "text_preview": "lock their vector dimensions"
}

Response:

{"id": 7, "status": "saved"}

`GET /ab-tests/{id}/highlights`

Get all highlights for a test.

Response:

{
  "test_id": "c9f3...",
  "highlights": [
    {
      "id": 7,
      "candidate": 1,
      "start_offset": 42,
      "end_offset": 67,
      "sentiment": "positive",
      "text_preview": "lock their vector dimensions"
    }
  ]
}

`POST /ab-scores`

Legacy flat-file A/B score submission. Writes scores plus hallucination check results to a JSON file in {CONTENT_PATH}/ab-scores/. Still functional; new work should use /ab-tests endpoints.

Frontend Features

Hash-based URL routing: #create, #drafts, #drafts/{path}, #ab-tests, #ab-tests/{id}
Breadcrumbs bar with back button support
Browser history integration

Create Flow — "Talk First, Classify After"

Record audio first, then pick content type from AI suggestions
Alternative: type or paste text, get suggestions
No content type selection required before recording
Classify session held server-side for 1 hour, consumed on format

Drafts View

Pipeline view with Untested and Tested sections
Per-draft actions: Copy to clipboard, Export as .md, Send to A/B Test
Drag-and-drop .txt/.md file import into the text input

A/B Test Review

Sequential review state machine: select → review_a → review_b → compare
Sidebar listing existing tests with status badges and score summaries
Auto-save every 120 seconds during review
Dimension rating 1-5 with comment fields per dimension

Keyboard Shortcuts

Key	Action
`1`–`5`	Rate focused dimension
`Tab` / `Shift+Tab`	Cycle through dimensions
`Enter`	Advance to next step
`Esc`	Navigate back

AI Scoring (Compare Step)

One-click AI scoring via GLM-4.7-Flash
15-second AbortController timeout with retry button on failure
5-column comparison table: dimension, human A rating, human B rating, AI A rating, AI B rating
Disagreement cells highlighted
Expandable rationale details per dimension

Inline Highlights

Select any phrase in either candidate to annotate as positive or negative
Floating toolbar on text selection
Highlights persisted to SQLite and re-rendered in compare view

Word-Level Diff

Toggle in compare view to see character-level differences between candidates
<ins>/<del> rendering via diff-match-patch (CDN)

Recording

16-bar waveform visualization via Web Audio API during recording

UX Polish

Skeleton loading states during async operations
Toast notification system: info, success, error — 5-second auto-dismiss
[EXPAND] markers in draft text rendered as styled callout boxes
Global error boundary

Output Format

Drafts are saved as markdown with YAML frontmatter:

---
type: til
date: 2026-03-15
tags: [qdrant, embeddings, debugging]
---

# embedding dimensions must match at collection creation time

Qdrant collections lock their vector dimensions at collection creation time.
If you ingest with 2560-dim vectors into a 1536-dim collection, every upsert
silently fails. Drop and recreate the collection to fix.

Filename format: YYYY-MM-DD-{slug}.md (slug derived from title, max 60 chars)

Written to: /mnt/apps/src/factory/content/drafts/{content_type}/

Dependencies

Service	Purpose	Required
haiven-transcribe	Speech-to-text (OpenAI-compatible `/v1/audio/transcriptions`)	Yes
litellm	LLM gateway — Hermes-4.3-36B for drafting, GLM-4.7-Flash for classify/AI scoring	Yes
traefik	HTTPS termination and domain routing	Yes

Technology Stack

Component	Technology
Runtime	Python 3.12
Web Framework	FastAPI + uvicorn
Database	SQLite via aiosqlite 0.20.0 (WAL mode, async)
LLM Client	openai SDK (LiteLLM-compatible)
Frontend	Vanilla HTML/CSS/JavaScript (no build step)
Frontend diff	diff-match-patch 1.0.5 (CDN)
Theme	Crimson Edge — dark warm-black (#0D0B0C), crimson accent (#E84444), mobile-first

Files

File	Purpose
`docker-compose.yml`	Service definition
`Dockerfile`	Container build
`requirements.txt`	Python dependencies
`SYSTEM_PROMPT.md`	Agent identity — content formatting rules and voice profiles
`app/main.py`	FastAPI application and all route definitions (672 lines)
`app/db.py`	Async SQLite database layer — 6 tables, WAL mode, CRUD (399 lines)
`app/migrate.py`	One-time JSON→SQLite data migration (180 lines)
`app/ai_score.py`	AI scoring and test name generation via GLM-4.7-Flash (187 lines)
`app/transcriber.py`	haiven-transcribe STT integration
`app/formatter.py`	LLM prompt construction, Hermes-4.3-36B calls, JSON parsing, content classification
`app/writer.py`	Markdown file writer with YAML frontmatter and slug generation
`app/prompts/content_types.py`	Content type definitions, voice registers, anti-pattern lists
`app/static/index.html`	Frontend HTML (~340 lines)
`app/static/style.css`	Frontend styles — Crimson Edge theme (~2,081 lines)
`app/static/app.js`	Frontend JavaScript — routing, A/B flow, AI scoring (~2,134 lines)
`app/static/favicon.svg`	Crimson square favicon
`data/factory.db`	SQLite database (runtime, not committed)

Source: /mnt/apps/docker/ai/content-factory/

Usage Examples

Health Check

curl -s https://factory.haiven.site/health | jq
# {"status": "healthy", "service": "content-factory"}

Classify Text, Then Format

# Step 1: classify
RESULT=$(curl -s -X POST https://factory.haiven.site/classify-text \
  -H "Content-Type: application/json" \
  -d '{"text": "qdrant silently drops upserts if vector dims do not match"}')
CLASSIFY_ID=$(echo $RESULT | jq -r '.classify_id')

# Step 2: format using the classify session
curl -X POST https://factory.haiven.site/format \
  -H "Content-Type: application/json" \
  -d "{\"classify_id\": \"$CLASSIFY_ID\", \"content_type\": \"til\"}" | jq

Format Text Directly

curl -X POST https://factory.haiven.site/format \
  -H "Content-Type: application/json" \
  -d '{
    "text": "qdrant silently drops upserts if vector dims do not match collection schema",
    "content_type": "til"
  }' | jq

Save a Draft

curl -X POST https://factory.haiven.site/save \
  -H "Content-Type: application/json" \
  -d '{"draft_id": "b4e2..."}' | jq

List Drafts

curl -s https://factory.haiven.site/drafts | jq

List A/B Tests Grouped by Voice Pair

curl -s "https://factory.haiven.site/ab-tests?group_by=voice_pair" | jq

Run AI Scoring on a Test

curl -X POST https://factory.haiven.site/ab-tests/c9f3.../ai-score | jq

Development

Rebuild After Source Changes

cd /mnt/apps/docker/ai/content-factory
docker compose build --no-cache && docker compose up -d

View Logs

docker compose logs -f content-factory

SQLite Database

The database lives at ./data/factory.db on the host (mounted to /data/factory.db in the container). WAL mode is enabled for concurrent reads. On first startup, migrate.py runs automatically and imports any existing flat-file A/B scores and hallucination data.

# Inspect the database
sqlite3 /mnt/apps/docker/ai/content-factory/data/factory.db

# Check tables
sqlite3 data/factory.db ".tables"

# Count A/B tests
sqlite3 data/factory.db "SELECT COUNT(*) FROM ab_tests WHERE status != 'archived';"

Security

HTTPS enforced via Traefik (required for browser microphone access)
No authentication — assumes trusted internal network
Audio not stored — transcribed in memory and immediately discarded
Input validation — file type and transcription content validated before LLM calls
Path traversal protection — GET /drafts/{path} resolves the path with os.path.realpath() and rejects any request that escapes CONTENT_PATH
Classify session TTL — classify sessions are pruned after 1 hour; they are also consumed (deleted) on successful /format call
Timestamp filenames — not user-controlled, preventing path traversal in draft writes

Monitoring

LLM calls are traced in Langfuse (LANGFUSE_ENABLED=true) via LiteLLM's callback integration. Each /format, /revise, and /ab-tests/{id}/ai-score call appears in Langfuse with content-factory and ai-score tags. View traces at https://ai-ops.haiven.site.

Troubleshooting

"Could not reach transcription service"

haiven-transcribe is not running or not reachable on the backend network.

docker ps | grep haiven-transcribe
docker exec content-factory curl -I http://haiven-transcribe:8000/health
docker logs haiven-transcribe | tail -30

"Could not reach LLM service"

LiteLLM or the backing model is unavailable.

docker ps | grep litellm
curl http://localhost:4000/v1/models | jq '.data[].id' | grep -E 'seed|glm'
docker logs litellm | tail -30

A/B test not found after restart

Draft sessions and A/B tests persist to SQLite (./data/factory.db). If the volume was not mounted, data will be lost. Verify the ./data:/data volume is present in the compose file.

Draft not saved — permission error

Check that the content directory is writable.

docker exec content-factory ls -la /content
ls -la /mnt/apps/src/factory/content
# directory must be owned by or writable by uid 1000
chown -R 1000:1000 /mnt/apps/src/factory/content

AI scoring returns empty or fails silently

Check that SCORING_MODEL (default glm-4-7-flash) is available in LiteLLM and that the model is currently loaded.

curl http://localhost:4000/v1/models | jq '.data[].id' | grep glm
docker logs content-factory | grep "AI scoring"

Microphone access denied in browser

HTTPS is required for browser microphone API. Access via https://factory.haiven.site only.

Service won't start

# Check port conflict
ss -tuln | grep 8023

# Full rebuild
docker compose down && docker compose up -d --build

Content Factory

Quick Start

Architecture

Components

Content Types

Voice Registers

Agent Identity

SQLite Schema

Service Configuration

Docker Compose

Environment Variables

Volumes

Networks

API Endpoints

GET /

GET /health

GET /types

POST /classify

POST /classify-text

POST /record

POST /format

POST /revise

POST /revise-text

POST /save

GET /drafts

GET /drafts/{path}

POST /ab-tests

GET /ab-tests

GET /ab-tests/{id}

POST /ab-tests/{id}/autosave

PATCH /ab-tests/{id}

GET /ab-tests/{id}/scores

DELETE /ab-tests/{id}

POST /ab-tests/{id}/ai-score

POST /ab-tests/generate-name

POST /ab-tests/{id}/highlights

GET /ab-tests/{id}/highlights

POST /ab-scores

Frontend Features

Navigation

Create Flow — "Talk First, Classify After"

Drafts View

A/B Test Review

Keyboard Shortcuts

AI Scoring (Compare Step)

Inline Highlights

Word-Level Diff

Recording

UX Polish

Output Format

Dependencies

Technology Stack

Files

Usage Examples

Health Check

Classify Text, Then Format

Format Text Directly

Save a Draft

List Drafts

List A/B Tests Grouped by Voice Pair

Run AI Scoring on a Test

Development

Rebuild After Source Changes

View Logs

SQLite Database

Security

Monitoring

Troubleshooting

"Could not reach transcription service"

"Could not reach LLM service"

A/B test not found after restart

Draft not saved — permission error

AI scoring returns empty or fails silently

Microphone access denied in browser

Service won't start

Links

`GET /`

`GET /health`

`GET /types`

`POST /classify`

`POST /classify-text`

`POST /record`

`POST /format`

`POST /revise`

`POST /revise-text`

`POST /save`

`GET /drafts`

`GET /drafts/{path}`

`POST /ab-tests`

`GET /ab-tests`

`GET /ab-tests/{id}`

`POST /ab-tests/{id}/autosave`

`PATCH /ab-tests/{id}`

`GET /ab-tests/{id}/scores`

`DELETE /ab-tests/{id}`

`POST /ab-tests/{id}/ai-score`

`POST /ab-tests/generate-name`

`POST /ab-tests/{id}/highlights`

`GET /ab-tests/{id}/highlights`

`POST /ab-scores`