PocketVeto/ARCHITECTURE.md

# PocketVeto — Architecture & Feature Documentation

> **App brand:** PocketVeto
> **Repo:** civicstack
> **Purpose:** Citizen-grade US Congress monitoring with AI-powered bill analysis, per-claim citations, and personalized tracking.

---

## Table of Contents

1. [Overview](#overview)
2. [Tech Stack](#tech-stack)
3. [Infrastructure & Docker](#infrastructure--docker)
4. [Configuration & Environment](#configuration--environment)
5. [Database Schema](#database-schema)
6. [Alembic Migrations](#alembic-migrations)
7. [Backend API](#backend-api)
8. [Celery Workers & Pipeline](#celery-workers--pipeline)
9. [LLM Service](#llm-service)
10. [Frontend](#frontend)
11. [Authentication](#authentication)
12. [Key Architectural Patterns](#key-architectural-patterns)
13. [Feature History](#feature-history)
14. [Deployment](#deployment)

---

## Overview

PocketVeto is a self-hosted, full-stack application that automatically tracks US Congress legislation, fetches bill text, generates AI summaries with per-claim source citations, correlates bills with news and Google Trends, and presents everything through a personalized dashboard. Users follow bills, members of Congress, and policy topics; the system surfaces relevant activity in their feed.

```
Congress.gov API → Poller → DB → Document Fetcher → GovInfo
                                                       ↓
                                                 LLM Processor
                                                       ↓
                                                  BillBrief
                                                 (cited AI brief)
                                                       ↓
                                              News Fetcher + Trend Scorer
                                                       ↓
                                               Next.js Frontend
```

---

## Tech Stack

| Layer | Technology |
|---|---|
| Reverse Proxy | Nginx (alpine) |
| Backend API | FastAPI + SQLAlchemy (async) |
| Task Queue | Celery 5 + Redis |
| Task Scheduler | Celery Beat + RedBeat (Redis-backed) |
| Database | PostgreSQL 16 |
| Cache / Broker | Redis 7 |
| Frontend | Next.js 15, React, Tailwind CSS, TypeScript |
| Auth | JWT (python-jose) + bcrypt (passlib) |
| LLM | Multi-provider factory: OpenAI, Anthropic, Gemini, Ollama |
| Bill Metadata | Congress.gov API (api.data.gov key) |
| Bill Text | GovInfo API (same api.data.gov key) |
| News | NewsAPI.org (100 req/day free tier) |
| Trends | Google Trends via pytrends |

---

## Infrastructure & Docker

### Services (`docker-compose.yml`)

```
postgres:16-alpine
  DB:       pocketveto
  User:     congress
  Port:     5432 (internal)

redis:7-alpine
  Port:     6379 (internal)
  Role:     Celery broker, result backend, RedBeat schedule store

api  (civicstack-api image)
  Port:     8000 (internal)
  Command:  alembic upgrade head && uvicorn app.main:app --host 0.0.0.0 --port 8000
  Depends:  postgres (healthy), redis (healthy)

worker  (civicstack-worker image)
  Command:  celery -A app.workers.celery_app worker -Q polling,documents,llm,news -c 4
  Depends:  postgres (healthy), redis (healthy)

beat  (civicstack-beat image)
  Command:  celery -A app.workers.celery_app beat -S redbeat.RedBeatScheduler
  Depends:  redis (healthy)

frontend  (civicstack-frontend image)
  Port:     3000 (internal)
  Build:    Next.js standalone output

nginx:alpine
  Port:     80 → public
  Routes:   /api/* → api:8000 | /* → frontend:3000
```

### Nginx Config (`nginx/nginx.conf`)

- `resolver 127.0.0.11 valid=10s` — re-resolves Docker DNS after container restarts (prevents stale-IP 502s on redeploy)
- `/api/` → FastAPI, 120s read timeout
- `/_next/static/` → frontend with 1-day cache header
- `/` → frontend with WebSocket upgrade support

---

## Configuration & Environment

Copy `.env.example` → `.env` and fill in keys before first run.

```env
# Network
LOCAL_URL=http://localhost
PUBLIC_URL=                          # optional, e.g. https://yourapp.com

# Auth
JWT_SECRET_KEY=                      # python -c "import secrets; print(secrets.token_hex(32))"

# PostgreSQL
POSTGRES_USER=congress
POSTGRES_PASSWORD=congress
POSTGRES_DB=pocketveto

# Redis
REDIS_URL=redis://redis:6379/0

# Congress.gov + GovInfo (shared key from api.data.gov)
DATA_GOV_API_KEY=
CONGRESS_POLL_INTERVAL_MINUTES=30

# LLM — pick one provider
LLM_PROVIDER=openai                  # openai | anthropic | gemini | ollama
OPENAI_API_KEY=
OPENAI_MODEL=gpt-4o
ANTHROPIC_API_KEY=
ANTHROPIC_MODEL=claude-opus-4-6
GEMINI_API_KEY=
GEMINI_MODEL=gemini-2.0-flash
OLLAMA_BASE_URL=http://host.docker.internal:11434
OLLAMA_MODEL=llama3.1

# News & Trends
NEWSAPI_KEY=
PYTRENDS_ENABLED=true
```

**Runtime overrides:** LLM provider/model and poll interval can be changed live through the Admin page — stored in the `app_settings` table and take precedence over env vars.

---

## Database Schema

### `bills`
Primary key: `bill_id` — natural key in format `{congress}-{type}-{number}` (e.g. `119-hr-1234`).

| Column | Type | Notes |
|---|---|---|
| bill_id | varchar (PK) | |
| congress_number | int | |
| bill_type | varchar | `hr`, `s`, `hjres`, `sjres` (tracked); `hres`, `sres`, `hconres`, `sconres` (not tracked) |
| bill_number | int | |
| title | text | |
| short_title | text | |
| sponsor_id | varchar (FK → members) | bioguide_id |
| introduced_date | date | |
| latest_action_date | date | |
| latest_action_text | text | |
| status | varchar | |
| chamber | varchar | House / Senate |
| congress_url | varchar | congress.gov link |
| govtrack_url | varchar | |
| last_checked_at | timestamptz | |
| actions_fetched_at | timestamptz | |
| created_at / updated_at | timestamptz | |

Indexes: `congress_number`, `latest_action_date`, `introduced_date`, `chamber`, `sponsor_id`

---

### `bill_actions`

| Column | Type | Notes |
|---|---|---|
| id | int (PK) | |
| bill_id | varchar (FK → bills, CASCADE) | |
| action_date | date | |
| action_text | text | |
| action_type | varchar | |
| chamber | varchar | |
| created_at | timestamptz | |

---

### `bill_documents`
Stores fetched bill text versions from GovInfo.

| Column | Type | Notes |
|---|---|---|
| id | int (PK) | |
| bill_id | varchar (FK → bills, CASCADE) | |
| doc_type | varchar | `bill_text`, `committee_report`, `amendment` |
| doc_version | varchar | Introduced, Enrolled, etc. |
| govinfo_url | varchar | Source URL on GovInfo |
| raw_text | text | Full extracted text |
| fetched_at | timestamptz | |
| created_at | timestamptz | |

---

### `bill_briefs`
AI-generated analysis. `key_points` and `risks` are JSONB arrays of cited objects.

| Column | Type | Notes |
|---|---|---|
| id | int (PK) | |
| bill_id | varchar (FK → bills, CASCADE) | |
| document_id | int (FK → bill_documents, SET NULL) | |
| brief_type | varchar | `full` (first version) or `amendment` (diff from prior version) |
| summary | text | 2-4 paragraph plain-language summary |
| key_points | jsonb | `[{text, citation, quote}]` |
| risks | jsonb | `[{text, citation, quote}]` |
| deadlines | jsonb | `[{date, description}]` |
| topic_tags | jsonb | `["healthcare", "taxation", ...]` |
| llm_provider | varchar | Which provider generated this brief |
| llm_model | varchar | Specific model name |
| govinfo_url | varchar (nullable) | Source document URL (from bill_documents) |
| created_at | timestamptz | |

Indexes: `bill_id`, `topic_tags` (GIN for JSONB containment queries)

**Citation structure** — each `key_points`/`risks` item:
```json
{
  "text": "The bill allocates $50B for defense",
  "citation": "Section 301(a)(2)",
  "quote": "There is hereby appropriated for fiscal year 2026, $50,000,000,000 for the Department of Defense...",
  "label": "cited_fact"
}
```
`label` is `"cited_fact"` when the claim is explicitly stated in the quoted text, or `"inference"` when it is an analytical interpretation. Old briefs without this field render without a badge (backward compatible).

---

### `members`
Primary key: `bioguide_id` (Congress.gov canonical identifier).

| Column | Type | Notes |
|---|---|---|
| bioguide_id | varchar (PK) | |
| name | varchar | Stored as "Last, First" |
| first_name / last_name | varchar | |
| party | varchar | |
| state | varchar | |
| chamber | varchar | |
| district | varchar (nullable) | House only |
| photo_url | varchar (nullable) | |
| official_url | varchar (nullable) | Member's official website |
| congress_url | varchar (nullable) | congress.gov profile link |
| birth_year | varchar(10) (nullable) | |
| address | varchar (nullable) | DC office address |
| phone | varchar(50) (nullable) | DC office phone |
| terms_json | json (nullable) | Array of `{congress, startYear, endYear, chamber, partyName, stateName, district}` |
| leadership_json | json (nullable) | Array of `{type, congress, current}` |
| sponsored_count | int (nullable) | Total bills sponsored (lifetime) |
| cosponsored_count | int (nullable) | Total bills cosponsored (lifetime) |
| detail_fetched | timestamptz (nullable) | Set when bio detail was enriched from Congress.gov |
| created_at / updated_at | timestamptz | |

Member detail fields (`congress_url` through `detail_fetched`) are populated lazily on first profile view via a Congress.gov detail API call. The `detail_fetched` timestamp is the gate for scheduling member interest scoring.

### `member_trend_scores`
One record per member per day (mirrors `trend_scores` for bills).

| Column | Type | Notes |
|---|---|---|
| id | int (PK) | |
| member_id | varchar (FK → members, CASCADE) | bioguide_id |
| score_date | date | |
| newsapi_count | int | Articles from NewsAPI (30-day window) |
| gnews_count | int | Articles from Google News RSS |
| gtrends_score | float | Google Trends interest 0–100 |
| composite_score | float | Weighted combination 0–100 (same formula as bill trend scores) |

Unique constraint: `(member_id, score_date)`. Indexes: `member_id`, `score_date`, `composite_score`.

### `member_news_articles`
News articles correlated to a specific member of Congress.

| Column | Type | Notes |
|---|---|---|
| id | int (PK) | |
| member_id | varchar (FK → members, CASCADE) | bioguide_id |
| source | varchar | News outlet |
| headline | text | |
| url | varchar | Unique per `(member_id, url)` |
| published_at | timestamptz | |
| relevance_score | float | Default 1.0 |
| created_at | timestamptz | |

---

### `users`

| Column | Type | Notes |
|---|---|---|
| id | int (PK) | |
| email | varchar (unique) | |
| hashed_password | varchar | bcrypt |
| is_admin | bool | First registered user = true |
| notification_prefs | jsonb | ntfy topic URL, ntfy auth token, ntfy enabled, RSS token |
| rss_token | varchar (nullable) | Unique token for personal RSS feed URL |
| created_at | timestamptz | |

---

### `follows`

| Column | Type | Notes |
|---|---|---|
| id | int (PK) | |
| user_id | int (FK → users, CASCADE) | |
| follow_type | varchar | `bill`, `member`, `topic` |
| follow_value | varchar | bill_id, bioguide_id, or topic name |
| follow_mode | varchar | `neutral` \| `pocket_veto` \| `pocket_boost` (default `neutral`) |
| created_at | timestamptz | |

Unique constraint: `(user_id, follow_type, follow_value)`

---

### `news_articles`

| Column | Type | Notes |
|---|---|---|
| id | int (PK) | |
| bill_id | varchar (FK → bills, CASCADE) | |
| source | varchar | News outlet |
| headline | varchar | |
| url | varchar | Unique per `(bill_id, url)` — same article can appear across multiple bills |
| published_at | timestamptz | |
| relevance_score | float | Default 1.0 |
| created_at | timestamptz | |

---

### `trend_scores`
One record per bill per day.

| Column | Type | Notes |
|---|---|---|
| id | int (PK) | |
| bill_id | varchar (FK → bills, CASCADE) | |
| score_date | date | |
| newsapi_count | int | Articles from NewsAPI (30-day window) |
| gnews_count | int | Articles from Google News RSS |
| gtrends_score | float | Google Trends interest 0–100 |
| composite_score | float | Weighted combination 0–100 |
| created_at | timestamptz | |

**Composite score formula:**
```
newsapi_pts  = min(newsapi_count / 20,  1.0) × 40   # saturates at 20 articles
gnews_pts    = min(gnews_count  / 50,  1.0) × 30   # saturates at 50 articles
gtrends_pts  = (gtrends_score / 100)        × 30
composite    = newsapi_pts + gnews_pts + gtrends_pts   # range 0–100
```

---

### `committees` / `committee_bills`

| committees | committee_id (PK), name, chamber, type |
|---|---|
| committee_bills | id, committee_id (FK), bill_id (FK), referred_date |

---

### `app_settings`
Key-value store for runtime-configurable settings.

| Key | Purpose |
|---|---|
| `congress_last_polled_at` | ISO timestamp of last successful poll |
| `llm_provider` | Overrides `LLM_PROVIDER` env var |
| `llm_model` | Overrides provider default model |
| `congress_poll_interval_minutes` | Overrides env var |

---

### `notifications`
Stores notification events for dispatching to user channels (ntfy, RSS).

| Column | Type | Notes |
|---|---|---|
| id | int (PK) | |
| user_id | int (FK → users, CASCADE) | |
| bill_id | varchar (FK → bills, SET NULL) | nullable |
| event_type | varchar | `new_document`, `new_amendment`, `bill_updated` |
| payload | jsonb | `{bill_title, bill_label, brief_summary, bill_url, milestone_tier}` |
| dispatched_at | timestamptz (nullable) | NULL = pending dispatch |
| created_at | timestamptz | |

`milestone_tier` in payload: `"progress"` (passed, signed, markup, conference, etc.) or `"referral"` (committee referral). Neutral follows silently skip referral-tier events; pocket_veto and pocket_boost receive them as early warnings.

---

## Alembic Migrations

| File | Description |
|---|---|
| `0001_initial_schema.py` | All initial tables |
| `0002_widen_chamber_party_columns.py` | Wider varchar for Bill.chamber, Member.party |
| `0003_widen_member_state_district.py` | Wider varchar for Member.state, Member.district |
| `0004_add_brief_type.py` | BillBrief.brief_type column (`full`/`amendment`) |
| `0005_add_users_and_user_follows.py` | users table + user_id FK on follows; drops global follows |
| `0006_add_brief_govinfo_url.py` | BillBrief.govinfo_url for frontend source links |
| `0007_add_member_bio_fields.py` | Member extended bio: `congress_url`, `birth_year`, `address`, `phone`, `terms_json`, `leadership_json`, `sponsored_count`, `cosponsored_count`, `detail_fetched` |
| `0008_add_member_interest_tables.py` | New tables: `member_trend_scores`, `member_news_articles` |
| `0009_fix_news_articles_url_uniqueness.py` | Changed `news_articles.url` from globally unique to per-bill unique `(bill_id, url)` |
| `0010_backfill_bill_congress_urls.py` | Backfill congress_url on existing bill records |
| `0011_add_notifications.py` | `notifications` table + `rss_token` column on users |

Migrations run automatically on API startup: `alembic upgrade head`.

---

## Backend API

Base URL: `/api`
Auth header: `Authorization: Bearer <jwt>`

### `/api/auth`

| Method | Path | Auth | Description |
|---|---|---|---|
| POST | `/register` | — | Create account. First user → admin. Returns token + user. |
| POST | `/login` | — | Returns token + user. |
| GET | `/me` | Required | Current user info. |

### `/api/bills`

| Method | Path | Auth | Description |
|---|---|---|---|
| GET | `/` | — | Paginated bill list. Query: `chamber`, `topic`, `sponsor_id`, `q`, `page`, `per_page`, `sort`. Includes `has_document` flag per bill via a single batch query. |
| GET | `/{bill_id}` | — | Full bill detail with sponsor, actions, briefs, news, trend scores. |
| GET | `/{bill_id}/actions` | — | Action timeline, newest first. |
| GET | `/{bill_id}/news` | — | Related news articles, limit 20. |
| GET | `/{bill_id}/trend` | — | Trend score history. Query: `days` (7–365, default 30). |
| POST | `/{bill_id}/draft-letter` | — | Generate a constituent letter draft via the configured LLM. Body: `{stance, recipient, tone, selected_points, include_citations, zip_code?}`. Returns `{draft: string}`. ZIP code is used in the prompt only — never stored or logged. |

### `/api/members`

| Method | Path | Auth | Description |
|---|---|---|---|
| GET | `/` | — | Paginated members. Query: `chamber`, `party`, `state`, `q`, `page`, `per_page`. |
| GET | `/{bioguide_id}` | — | Member detail. On first view, lazily enriches bio from Congress.gov and queues member interest scoring. Returns `latest_trend` if scored. |
| GET | `/{bioguide_id}/bills` | — | Member's sponsored bills, paginated. |
| GET | `/{bioguide_id}/trend` | — | Member trend score history. Query: `days` (7–365, default 30). |
| GET | `/{bioguide_id}/news` | — | Member's recent news articles, limit 20. |

### `/api/follows`

| Method | Path | Auth | Description |
|---|---|---|---|
| GET | `/` | Required | Current user's follows. |
| POST | `/` | Required | Add follow `{follow_type, follow_value}`. Idempotent. |
| DELETE | `/{id}` | Required | Remove follow (ownership checked). |

### `/api/dashboard`

| Method | Path | Auth | Description |
|---|---|---|---|
| GET | `/` | Required | Personalized feed from followed bills/members/topics + trending. Returns `{feed, trending, follows}`. |

### `/api/search`

| Method | Path | Auth | Description |
|---|---|---|---|
| GET | `/` | — | Full-text search. Query: `q` (min 2 chars). Returns `{bills, members}`. |

### `/api/settings`

| Method | Path | Auth | Description |
|---|---|---|---|
| GET | `/` | Required | Current settings (DB overrides env). |
| PUT | `/` | Admin | Update `{key, value}`. Allowed keys: `llm_provider`, `llm_model`, `congress_poll_interval_minutes`. |
| POST | `/test-llm` | Admin | Test LLM connection with a lightweight ping (max_tokens=20). Returns `{status, provider, model, reply}`. |
| GET | `/llm-models?provider=X` | Admin | Fetch available models from the live provider API. Supports openai, anthropic, gemini, ollama. |

### `/api/notifications`

| Method | Path | Auth | Description |
|---|---|---|---|
| GET | `/settings` | Required | User's notification preferences (ntfy URL/token, ntfy enabled, RSS token). |
| PUT | `/settings` | Required | Update notification preferences. |
| POST | `/settings/rss-reset` | Required | Regenerate RSS token (invalidates old URL). |
| GET | `/feed/{rss_token}.xml` | — | Personal RSS feed of notification events for this user. |

### `/api/admin`

| Method | Path | Auth | Description |
|---|---|---|---|
| GET | `/users` | Admin | All users with follow counts. |
| DELETE | `/users/{id}` | Admin | Delete user (cannot delete self). Cascades follows. |
| PATCH | `/users/{id}/toggle-admin` | Admin | Promote/demote admin status (cannot change self). |
| GET | `/stats` | Admin | Pipeline counters: total bills, docs fetched, briefs generated, pending LLM, missing metadata/sponsors/actions, uncited briefs, unlabeled briefs (cited objects without a fact/inference label). |
| GET | `/api-health` | Admin | Test each external API in parallel; returns status + latency for Congress.gov, GovInfo, NewsAPI, Google News. |
| POST | `/trigger-poll` | Admin | Queue immediate Congress.gov poll. |
| POST | `/trigger-member-sync` | Admin | Queue member sync. |
| POST | `/trigger-trend-scores` | Admin | Queue trend score calculation. |
| POST | `/trigger-fetch-actions` | Admin | Queue action fetch for recently active bills (last 30 days). |
| POST | `/backfill-all-actions` | Admin | Queue action fetch for ALL bills with no action history (one-time catch-up). |
| POST | `/backfill-sponsors` | Admin | Queue one-off task to populate `sponsor_id` on bills where it is NULL. |
| POST | `/backfill-metadata` | Admin | Fill null `introduced_date`, `chamber`, `congress_url` by re-fetching bill detail. |
| POST | `/backfill-citations` | Admin | Delete pre-citation briefs and re-queue LLM using stored document text. |
| POST | `/backfill-labels` | Admin | Classify existing cited brief points as `cited_fact` or `inference` in-place — one compact LLM call per brief, no re-generation. |
| POST | `/resume-analysis` | Admin | Re-queue LLM for docs with no brief; re-queue doc fetch for bills with no doc. |
| POST | `/bills/{bill_id}/reprocess` | Admin | Queue document + action fetches for a specific bill (debugging). |
| GET | `/task-status/{task_id}` | Admin | Celery task status and result. |

### `/api/health`

| Method | Path | Description |
|---|---|---|
| GET | `/` | Simple health check `{status: "ok", timestamp}`. |
| GET | `/detailed` | Tests PostgreSQL + Redis. Returns per-service status. |

---

## Celery Workers & Pipeline

**Celery app name:** `pocketveto`
**Broker / Backend:** Redis

### Queue Routing

| Queue | Workers | Tasks |
|---|---|---|
| `polling` | worker | `app.workers.congress_poller.*`, `app.workers.notification_dispatcher.*` |
| `documents` | worker | `fetch_bill_documents` |
| `llm` | worker | `process_document_with_llm` |
| `news` | worker | `app.workers.news_fetcher.*`, `app.workers.trend_scorer.*`, `app.workers.member_interest.*` |

**Worker settings:**
- `task_acks_late = True` — task removed from queue only after completion, not on pickup
- `worker_prefetch_multiplier = 1` — prevents workers from hoarding LLM tasks
- Serialization: JSON

### Beat Schedule (RedBeat, stored in Redis)

| Schedule | Task | When |
|---|---|---|
| Configurable (default 30 min) | `poll_congress_bills` | Continuous |
| Every 6 hours | `fetch_news_for_active_bills` | Ongoing |
| Daily 2 AM UTC | `calculate_all_trend_scores` | Nightly |
| Every 12 hours (at :30) | `fetch_news_for_active_members` | Ongoing |
| Daily 3 AM UTC | `calculate_all_member_trend_scores` | Nightly |
| Daily 4 AM UTC | `fetch_actions_for_active_bills` | Nightly |
| Every 5 minutes | `dispatch_notifications` | Continuous |

---

### Pipeline Flow

```
1. congress_poller.poll_congress_bills()
   ↳ Fetches bills updated since last poll (fromDateTime param)
   ↳ Filters: only hr, s, hjres, sjres (legislation that can become law)
   ↳ First run: seeds from 60 days back
   ↳ New bills: fetches bill detail endpoint to get sponsor (list endpoint
     has no sponsor data), upserts Member, sets bill.sponsor_id
   ↳ New bills → fetch_bill_documents.delay(bill_id)
   ↳ Updated bills → fetch_bill_documents.delay(bill_id) if changed
   ↳ Updated bills → emit bill_updated notification if action is a milestone:
       - "progress" tier: passed/failed, signed/vetoed, enacted, markup, conference,
         reported from committee, placed on calendar, cloture, roll call
         → all follow types (bill, sponsor, topic) receive notification
       - "referral" tier: referred to committee
         → pocket_veto and pocket_boost only; neutral follows silently skip

2. document_fetcher.fetch_bill_documents(bill_id)
   ↳ Gets text versions from Congress.gov (XML preferred, falls back to HTML/PDF)
   ↳ Fetches raw text from GovInfo
   ↳ Idempotent: skips if doc_version already stored
   ↳ Stores BillDocument with govinfo_url + raw_text
   ↳ → process_document_with_llm.delay(document_id)

3. llm_processor.process_document_with_llm(document_id)
   ↳ Rate limited: 10/minute
   ↳ Idempotent: skips if brief exists for document
   ↳ Determines type:
       - No prior brief → "full" brief
       - Prior brief exists → "amendment" brief (diff vs previous)
   ↳ Calls configured LLM provider
   ↳ Stores BillBrief with cited key_points and risks
   ↳ → fetch_news_for_bill.delay(bill_id)

4. news_fetcher.fetch_news_for_bill(bill_id)
   ↳ Queries NewsAPI + Google News RSS using bill title/number
   ↳ Deduplicates by (bill_id, url) — same article can appear for multiple bills
   ↳ Stores NewsArticle records

5. trend_scorer.calculate_all_trend_scores()  [nightly]
   ↳ Bills active in last 90 days
   ↳ Skips bills already scored today
   ↳ Fetches: NewsAPI count + Google News RSS count + Google Trends score
   ↳ Calculates composite_score (0–100)
   ↳ Stores TrendScore record

Member interest pipeline (independent of bill pipeline):

6. member_interest.fetch_member_news(bioguide_id)  [on first profile view + every 12h]
   ↳ Triggered on first member profile view (non-blocking via .delay())
   ↳ Queries NewsAPI + Google News RSS using member name + title
   ↳ Deduplicates by (member_id, url)
   ↳ Stores MemberNewsArticle records

7. member_interest.calculate_member_trend_score(bioguide_id)  [on first profile view + nightly]
   ↳ Triggered on first member profile view (non-blocking via .delay())
   ↳ Only runs if member detail has been fetched (gate: detail_fetched IS NOT NULL)
   ↳ Fetches: NewsAPI count + Google News RSS count + Google Trends score
   ↳ Uses the same composite formula as bills
   ↳ Stores MemberTrendScore record
```

---

## LLM Service

**File:** `backend/app/services/llm_service.py`

### Provider Factory

```python
get_llm_provider() → LLMProvider
```

Reads `LLM_PROVIDER` from AppSetting (DB) then env var. Instantiates the matching provider class.

| Provider | Class | Key Setting |
|---|---|---|
| `openai` | `OpenAIProvider` | `OPENAI_API_KEY`, `OPENAI_MODEL` |
| `anthropic` | `AnthropicProvider` | `ANTHROPIC_API_KEY`, `ANTHROPIC_MODEL` |
| `gemini` | `GeminiProvider` | `GEMINI_API_KEY`, `GEMINI_MODEL` |
| `ollama` | `OllamaProvider` | `OLLAMA_BASE_URL`, `OLLAMA_MODEL` |

All providers implement:
```python
generate_brief(doc_text, bill_metadata) → ReverseBrief
generate_amendment_brief(new_text, prev_text, bill_metadata) → ReverseBrief
generate_text(prompt) → str   # free-form text, used by draft letter generator
```

### ReverseBrief Dataclass

```python
@dataclass
class ReverseBrief:
    summary: str
    key_points: list[dict]   # [{text, citation, quote, label}]
    risks: list[dict]        # [{text, citation, quote, label}]
    deadlines: list[dict]    # [{date, description}]
    topic_tags: list[str]
    llm_provider: str
    llm_model: str
```

### Prompt Design

**Full brief prompt** instructs the LLM to produce:
```json
{
  "summary": "2-4 paragraph plain-language explanation",
  "key_points": [
    {
      "text": "claim",
      "citation": "Section X(y)",
      "quote": "verbatim excerpt ≤80 words",
      "label": "cited_fact"
    }
  ],
  "risks": [
    {
      "text": "concern",
      "citation": "Section X(y)",
      "quote": "verbatim excerpt ≤80 words",
      "label": "inference"
    }
  ],
  "deadlines": [{"date": "YYYY-MM-DD or null", "description": "..."}],
  "topic_tags": ["healthcare", "taxation"]
}
```

`label` classification rules baked into the system prompt: `"cited_fact"` if the claim is explicitly stated in the quoted text; `"inference"` if it is an analytical interpretation, projection, or implication not literally stated. The UI shows a neutral "Inferred" badge on inference items only (cited_fact is the clean default).

**Amendment brief prompt** focuses on what changed between document versions.

**Smart truncation:** Bills exceeding the token budget are trimmed — 75% of budget from the start (preamble/purpose), 25% from the end (enforcement/effective dates), with an omission notice in the middle.

**Token budgets:**
- OpenAI / Anthropic / Gemini: 6,000 tokens
- Ollama: 3,000 tokens (local models have smaller context windows)

---

## Frontend

**Framework:** Next.js 15 (App Router), TypeScript, Tailwind CSS
**State:** Zustand (auth), TanStack Query (server state)
**HTTP:** Axios with JWT interceptor

### Pages

| Route | Description |
|---|---|
| `/` | Dashboard — personalized feed + trending bills |
| `/bills` | Browse all bills with search, chamber/topic filters, pagination |
| `/bills/[id]` | Bill detail — brief with § citations, action timeline, news, trend chart |
| `/members` | Browse members of Congress, filter by chamber/party/state |
| `/members/[id]` | Member profile — bio, contact info, leadership roles, service history, sponsored bills, public interest trend chart, recent news |
| `/following` | User's followed bills, members, and topics |
| `/topics` | Browse and follow policy topics |
| `/settings` | Admin panel (admin only) |
| `/login` | Email + password sign-in |
| `/register` | Account creation |

### Key Components

**`BriefPanel.tsx`**
Orchestrates AI brief display. If the latest brief is type `amendment`, shows an amber "What Changed" badge. Renders the latest brief via `AIBriefCard`. Below it, a collapsible "Version History" lists all older briefs; clicking one expands an inline `AIBriefCard`.

**`AIBriefCard.tsx`**
Renders the LLM brief. For cited items (new format), shows a `§ Section X(y)` chip next to each bullet. Clicking the chip expands an inline panel with:
- Blockquoted verbatim excerpt from the bill
- "View source →" link to GovInfo (opens in new tab)
- One chip open at a time per card
- Inference items show a neutral "Inferred" badge (analytical interpretation, not a literal quote)
- Old plain-string briefs render without chips (graceful backward compat)

**`ActionTimeline.tsx`**
Renders the legislative action history as a vertical timeline. Accepts optional `latestActionDate`/`latestActionText` fallback props — when `actions` is empty but a latest action exists (actions not yet fetched from Congress.gov), shows a single "latest known action" entry with a note that full history loads in the background.

**`MobileHeader.tsx`**
Top bar shown only on mobile (`md:hidden`). Displays the PocketVeto logo and a hamburger button that opens the slide-in drawer.

**`AuthGuard.tsx`**
Client component wrapping the entire app. Waits for Zustand hydration, then redirects unauthenticated users to `/login`. Public paths (`/login`, `/register`) bypass the guard. Implements the responsive shell: desktop sidebar always-visible (`hidden md:flex`), mobile drawer with backdrop overlay controlled by `drawerOpen` state.

**`Sidebar.tsx`**
Navigation with: Home, Bills, Members, Following, Topics, Settings (admin only). Shows current user email + logout button at the bottom. Accepts optional `onClose` prop — when provided (mobile drawer context), renders an X close button in the header and calls `onClose` on every nav link click.

**`DraftLetterPanel.tsx`**
Collapsible panel rendered below `BriefPanel` on the bill detail page (only when a brief exists). Lets users select up to 3 cited points from the brief, choose stance (YES/NO), tone (short/polite/firm), and optionally enter a ZIP code (not stored). Stance auto-populates from the user's follow mode (`pocket_boost` → YES, `pocket_veto` → NO); clears if they unfollow. Recipient (house/senate) is derived from the bill's chamber. Calls `POST /{bill_id}/draft-letter` and renders the plain-text draft in a readonly textarea with a copy-to-clipboard button.

**`BillCard.tsx`**
Compact bill preview showing bill ID, title, sponsor with party badge, latest action date, status, and a text availability indicator: `Brief` (green, analysis done) / `Pending` (amber, text retrieved but not yet analysed) / `No text` (muted, nothing published on Congress.gov).

**`TrendChart.tsx`**
Line chart of `composite_score` over time with tooltip breakdown of each data source.

**`WelcomeBanner.tsx`**
Dismissible onboarding card rendered at the top of the dashboard. Shown only to guests (no JWT token). On dismiss — via the × button, the "Dismiss" link, or the "Browse Bills" CTA — sets `localStorage["pv_seen_welcome"] = "1"` and hides permanently. Reads localStorage after mount to avoid hydration mismatch; renders nothing until client-side state is resolved.

### Utility Functions (`lib/utils.ts`)

```typescript
partyBadgeColor(party) → Tailwind classes
  "Republican" → "bg-red-600 text-white"
  "Democrat"   → "bg-blue-600 text-white"
  other        → "bg-slate-500 text-white"

chamberBadgeColor(chamber) → Tailwind badge classes
  "Senate"  → amber/gold  (bg-amber-100 text-amber-700 …)
  "House"   → slate/silver (bg-slate-100 text-slate-600 …)

partyColor(party) → text color class (used inline)
trendColor(score) → color class based on score thresholds
billLabel(type, number) → "H.R. 1234", "S. 567", etc.
formatDate(date) → "Feb 28, 2026"
```

### Auth Store (`stores/authStore.ts`)

```typescript
interface AuthState {
  token: string | null
  user: { id: number; email: string; is_admin: boolean } | null
  setAuth(token, user): void
  logout(): void
}
// Persisted to localStorage as "pocketveto-auth"
```

**Auth-aware query keys:** TanStack Query keys that return different data for guests vs authenticated users include `!!token` in their key (e.g. `["dashboard", !!token]`). This ensures a fresh fetch fires automatically on login or logout without manual cache invalidation.

---

## Authentication

- **Algorithm:** HS256 JWT, 7-day expiry
- **Storage:** Zustand store persisted to `localStorage` key `pocketveto-auth`
- **Injection:** Axios request interceptor reads from localStorage and adds `Authorization: Bearer <token>` to every request
- **First user:** The first account registered automatically receives `is_admin = true`
- **Admin role:** Required for PUT/POST `/api/settings`, all `/api/admin/*` endpoints, and viewing the Settings page in the UI
- **No email verification:** Accounts are active immediately on registration
- **Public endpoints:** `/api/bills`, `/api/members`, `/api/search`, `/api/health` — no auth required

---

## Key Architectural Patterns

### Idempotent Workers
Every Celery task checks for existing records before processing. Combined with `task_acks_late=True`, this means:
- Tasks can be retried without creating duplicates
- Worker crashes don't lose work (task stays in queue until acknowledged)

### Incremental Polling
The Congress.gov poller uses `fromDateTime` to fetch only recently updated bills, tracking the last poll timestamp in `app_settings`. On first run it seeds 60 days back to avoid processing thousands of old bills.

### Bill Type Filtering
Only tracks legislation that can become law:
- `hr` (House Resolution → Bill)
- `s` (Senate Bill)
- `hjres` (House Joint Resolution)
- `sjres` (Senate Joint Resolution)

Excluded (procedural, cannot become law): `hres`, `sres`, `hconres`, `sconres`

### Queue Specialization
Separate queues prevent a flood of LLM tasks from blocking time-sensitive polling tasks. Worker prefetch of 1 prevents any single worker from hoarding slow LLM jobs.

### LLM Provider Abstraction
All LLM providers implement the same interface. Switching providers is a single admin setting change — no code changes, no restart required (the factory reads from DB on each task invocation).

### JSONB for Flexible Brief Storage
`key_points`, `risks`, `deadlines`, `topic_tags` are stored as JSONB. This means schema changes (adding `citation`/`quote` in v0.2.0, adding `label` in v0.6.0) required no migrations — only the LLM prompt and application code changed. Old string-format briefs, cited-object briefs without labels, and fully-labelled briefs all coexist in the same column and render correctly at each fidelity level.

### Redis-backed Beat Schedule (RedBeat)
The Celery Beat schedule is stored in Redis rather than in memory. This means the beat scheduler can restart without losing schedule state or double-firing tasks.

### Docker DNS Re-resolution
Nginx uses `resolver 127.0.0.11 valid=10s` (Docker's internal DNS) so upstream container IPs are refreshed every 10 seconds. Without this, nginx caches the IP at startup and returns 502 errors after any container is recreated.

---

## Feature History

### v0.1.0 — Foundation
- Docker Compose stack: PostgreSQL, Redis, FastAPI, Celery, Next.js, Nginx
- Congress.gov API integration: bill polling, member sync
- GovInfo document fetching with intelligent truncation
- Multi-provider LLM service (OpenAI, Anthropic, Gemini, Ollama)
- AI brief generation: summary, key points, risks, deadlines, topic tags
- Amendment-aware processing: diffs new bill versions against prior
- NewsAPI + Google News RSS article correlation
- Google Trends (pytrends) scoring
- Composite trend score (0–100) with weighted formula
- Full-text bill search (PostgreSQL tsvector)
- Member of Congress browsing
- Global follows (bill / member / topic)
- Personalized dashboard feed
- Admin settings page (LLM provider selection, data source status)
- Manual Celery task triggers from UI
- Bill type filtering: only legislation that can become law
- 60-day seed window on fresh install

**Multi-User Auth (added to v0.1.0):**
- Email + password registration/login (JWT, bcrypt)
- Per-user follow scoping
- Admin role (first user = admin)
- Admin user management: list, delete, promote/demote
- AuthGuard with login/register pages
- Analysis status dashboard (auto-refresh every 30s)

### v0.3.0 — Member Profiles & Mobile UI

**Member Interest Tracking:**
- `member_trend_scores` and `member_news_articles` tables (migration 0008)
- `member_interest` Celery worker: `fetch_member_news`, `calculate_member_trend_score`, `fetch_news_for_active_members`, `calculate_all_member_trend_scores`
- Member interest scoring uses the identical composite formula as bills (NewsAPI + GNews + pytrends)
- New beat schedules: member news every 12h, member trend scores nightly at 3 AM UTC
- Lazy enrichment: on first profile view, bio is fetched from Congress.gov detail API and interest scoring is queued non-blocking
- Member detail fields added: `congress_url`, `birth_year`, `address`, `phone`, `terms_json`, `leadership_json`, `sponsored_count`, `cosponsored_count`, `detail_fetched` (migration 0007)
- New API endpoints: `GET /api/members/{id}/trend` and `GET /api/members/{id}/news`
- Member detail page redesigned: photo, bio header with party/state/district/birth year, contact info (address, phone, website, congress.gov), current leadership badges, trend chart ("Public Interest"), news panel, legislation stats (sponsored/cosponsored counts), full service history timeline, all leadership roles history

**News Deduplication Fix:**
- `news_articles.url` changed from globally unique to per-bill unique `(bill_id, url)` (migration 0009)
- The same article can now appear in multiple bills' news panels
- `fetch_news_for_bill` now fetches from both NewsAPI and Google News RSS (previously GNews was volume-signal only)

**Mobile UI:**
- `MobileHeader.tsx` — hamburger + logo top bar, hidden on desktop (`md:hidden`)
- `AuthGuard.tsx` — responsive shell: desktop sidebar always-on, mobile slide-in drawer with backdrop
- `Sidebar.tsx` — `onClose` prop for drawer mode (X button + close on nav click)
- Dashboard grid: `grid-cols-1 md:grid-cols-3` (single column on mobile)
- Members page: `grid-cols-1 sm:grid-cols-2` (single column on mobile, two on tablet+)
- Topics page: `grid-cols-1 sm:grid-cols-2`

### v0.4.0 — Notifications, Admin Health Panel, Bill Action Pipeline

**Notifications (Phase 1 complete):**
- `notifications` table — stores events per user (new_brief, bill_updated, new_action)
- ntfy dispatch — Celery task POSTs to user's ntfy topic URL (self-hosted or ntfy.sh); optional auth token
- RSS feed — tokenized per-user XML feed at `/api/notifications/feed/{token}.xml`
- `dispatch_notifications` beat task — runs every 5 minutes, fans out unsent events to enabled channels
- Notification settings UI — ntfy topic URL, auth token, enable/disable, RSS URL with copy button

**Bill Action Pipeline:**
- `fetch_bill_actions` Celery task — fetches full legislative history from Congress.gov, idempotent on `(bill_id, action_date, action_text)`, updates `Bill.actions_fetched_at`
- `fetch_actions_for_active_bills` nightly batch — queues action fetches for bills active in last 30 days
- `backfill_all_bill_actions` — one-time task to fetch actions for all bills with `actions_fetched_at IS NULL`
- Beat schedule entry at 4 AM UTC
- `ActionTimeline` updated: shows full history when fetched; falls back to `latest_action_date`/`latest_action_text` with "latest known action" label when history not yet loaded

**"What Changed" — BriefPanel:**
- New `BriefPanel.tsx` component wrapping `AIBriefCard`
- When latest brief is type `amendment`: shows amber "What Changed" badge row + date
- Collapsible "Version History" section listing older briefs (date, type badge, truncated summary)
- Clicking a history row expands an inline `AIBriefCard` for that version

**LLM Provider Improvements:**
- Live model picker — `GET /api/settings/llm-models?provider=X` fetches available models from each provider's API (OpenAI SDK, Anthropic REST, Gemini SDK, Ollama tags endpoint)
- DB overrides now fully propagated: `get_llm_provider(provider, model)` accepts explicit params; all call sites read from `app_settings`
- Default Gemini model updated: `gemini-1.5-pro` (deprecated) → `gemini-2.0-flash`
- Test connection replaced with lightweight ping (max_tokens=20, 3-word prompt) instead of full brief generation

**Admin Panel Overhaul:**
- Bill Pipeline section: progress bar + breakdown table (total, text published, no text yet, AI briefs, pending LLM, uncited)
- External API Health: Run Tests button, parallel health checks for Congress.gov / GovInfo / NewsAPI / Google News RSS with latency display
- Manual Controls redesigned as health panel: each action has a status dot (green/red/gray), description, contextual count badge (e.g. "⚠ 12 bills missing metadata"), and Run button
- Task status polling: after triggering a task, button shows spinning icon; polls `/api/admin/task-status/{id}` every 5s; shows task ID prefix + completion/failure state
- New stat fields: `bills_missing_sponsor`, `bills_missing_metadata`, `bills_missing_actions`, `pending_llm`, `no_text_bills`
- New admin tasks: Backfill Dates & Links, Backfill All Action Histories, Resume Analysis

**Chamber Color Badges:**
- `chamberBadgeColor(chamber)` utility: amber/gold for Senate, slate/silver for House
- Applied everywhere chamber is displayed: BillCard, bill detail header

**Bill Detail Page:**
- "No bill text published" state — shown when `has_document=false` and no briefs; includes bill label, date, and congress.gov link
- `has_document` field added to `BillDetailSchema` and `BillDetail` TypeScript type
- `introduced_date` shown conditionally (not rendered when null, preventing "Introduced: —")
- Admin reprocess endpoint: `POST /api/admin/bills/{bill_id}/reprocess`

### v0.5.0 — Follow Modes, Public Browsing & Draft Letter Generator

**Follow Modes:**
- `follow_mode` column on `follows` table: `neutral | pocket_veto | pocket_boost`
- `FollowButton` replaced with a mode-selector dropdown (shield/zap/heart icons, descriptions for each mode)
- `pocket_veto` — alert only on advancement milestones; `pocket_boost` — all changes + action prompts
- Mode stored per-follow; respected by notification dispatcher

**Public Browsing:**
- Unauthenticated guests can browse bills, members, topics, and the trending dashboard
- `AuthModal` gates follow and other interactive actions
- Sidebar and nav adapt to guest state (no email/logout shown)
- All public endpoints already auth-free; guard refactored to allow guest reads

**Draft Constituent Letter Generator (email_gen):**
- `DraftLetterPanel.tsx` — collapsible UI below `BriefPanel` for bills with a brief
- User selects up to 3 cited points from the brief, picks stance (YES/NO), tone, optional ZIP (not stored)
- Stance pre-fills from follow mode; clears on unfollow (ref-tracked, not effect-guarded)
- Recipient derived from bill chamber — no dropdown needed
- `POST /api/bills/{bill_id}/draft-letter` endpoint: reads LLM provider/model from `AppSetting` (respects Settings page), wraps LLM errors with human-readable messages (quota, rate limit, auth)
- `generate_text(prompt) → str` added to `LLMProvider` ABC and all four providers

**Bill Text Status Indicators:**
- `has_document` field added to `BillSchema` (list endpoint) via a single batch `SELECT DISTINCT` — no per-card queries
- `BillCard` shows: `Brief` (green) / `Pending` (amber) / `No text` (muted) based on brief + document state

### v0.6.0 — Phase 2: Change-driven Alerts & Fact/Inference Labeling

**Change-driven Alerts:**
- `notification_utils.py` milestone keyword list expanded: added `"markup"` (markup sessions) and `"conference"` (conference committee)
- New `is_referral_action()` classifier for committee referrals (`"referred to"`)
- Two-tier notification system: `milestone_tier` field in `NotificationEvent.payload`
  - `"progress"` — high-signal milestones (passed, signed, markup, etc.): all follow types notified
  - `"referral"` — committee referral: pocket_veto and pocket_boost notified; neutral silently dropped
- **Topic followers now receive `bill_updated` milestone notifications** — previously they only received `new_document`/`new_amendment` events. Fixed by querying the bill's latest brief for `topic_tags` inside `_update_bill_if_changed()`
- All three follow types (bill, sponsor, topic) covered for both tiers

**Fact vs Inference Labeling:**
- `label: "cited_fact" | "inference"` added to every cited key_point and risk in the LLM JSON schema
- System prompt updated for all four providers (OpenAI, Anthropic, Gemini, Ollama)
- UI: neutral "Inferred" badge shown next to inference items in `AIBriefCard`; cited_fact items render cleanly without a badge
- `backfill_brief_labels` Celery task: classifies existing cited points in-place — one compact LLM call per brief (all points batched), updates JSONB with `flag_modified`, no brief re-generation
- `POST /api/admin/backfill-labels` endpoint + "Backfill Fact/Inference Labels" button in Admin panel
- `unlabeled_briefs` counter added to `/api/admin/stats` and pipeline breakdown table

**Admin Panel Cleanup:**
- Manual Controls split into two sections: always-visible recurring controls (Poll, Members, Trends, Actions, Resume) and a collapsible **Maintenance** section for one-time backfill tasks
- Maintenance section header shows "⚠ action needed" when any backfill has a non-zero count

### v0.6.1 — Welcome Banner & Dashboard Auth Fix

**Welcome Banner:**
- `WelcomeBanner.tsx` — dismissible onboarding card shown only to guests at the top of the dashboard
- Three bullet points: follow bills/members/topics, see what changed, Back to Source citations
- "Browse Bills" CTA navigates to `/bills` and dismisses; × and "Dismiss" button also dismiss
- Dismissed state stored in `localStorage["pv_seen_welcome"]`; never shown to logged-in users

**Dashboard Auth-Aware Query Key:**
- `useDashboard` hook query key changed from `["dashboard"]` to `["dashboard", !!token]`
- Fixes stale cache issue where logging in showed the guest feed until a manual refresh

### v0.2.2 — Sponsor Linking & Search Fixes
- **Root cause fixed:** Congress.gov list API does not return sponsor data — only the detail endpoint does. Poller now calls the detail endpoint for each new bill to get the sponsor and populate `bill.sponsor_id`
- **Backfill task:** `backfill_sponsor_ids` Celery task + `/api/admin/backfill-sponsors` endpoint + "Backfill Sponsors" button in Admin UI — fixes existing bills with `NULL` sponsor_id (~10 req/sec, ~3 min for 1,600 bills)
- **Member name search:** members are stored as "Last, First" in the `name` column; search now also matches "First Last" order using PostgreSQL `split_part()` — applied to both the Members page and global search
- **Search spaces:** removed `.trim()` on search `onChange` handlers in Members and Bills pages that was eating spaces as you typed
- **Member bills 500 error:** `get_member_bills` endpoint now eagerly loads `Bill.sponsor` via `selectinload` to prevent Pydantic MissingGreenlet error during serialization

### v0.2.0 — Citations
- **Per-claim citations on AI briefs:** every key point and risk includes:
  - `citation` — section reference (e.g., "Section 301(a)(2)")
  - `quote` — verbatim excerpt ≤80 words from that section
- `§` citation chip UI on each bullet — click to expand quote + GovInfo source link
- `govinfo_url` stored on `BillBrief` for direct frontend access
- Old briefs (plain strings) render without chips — backward compatible
- Migration 0006: `govinfo_url` column on `bill_briefs`
- Party badges redesigned: solid `red-600` / `blue-600` / `slate-500` with white text, readable in both light and dark mode
- Tailwind content scan extended to include `lib/` directory
- Nginx DNS resolver fix: prevents stale-IP 502s after container restarts

---

## Deployment

### First Deploy

```bash
cp .env.example .env
# Edit .env — add API keys, generate JWT_SECRET_KEY
docker compose up --build -d
```

Migrations run automatically. Navigate to the app, register the first account (it becomes admin).

### Updating

```bash
git pull origin main
docker compose up --build -d
docker compose exec nginx nginx -s reload   # if nginx wasn't recreated
```

### Useful Commands

```bash
# Check all service status
docker compose ps

# View logs
docker compose logs api --tail=50
docker compose logs worker --tail=50

# Force a bill poll now
# → Admin page → Manual Controls → Trigger Poll

# Check DB column layout
docker compose exec postgres psql -U congress -d pocketveto -c "\d bill_briefs"

# Tail live worker output
docker compose logs -f worker

# Restart a specific service
docker compose restart worker
```

### Bill Regeneration (Optional)

Existing briefs generated before v0.2.0 use plain strings (no citations). To regenerate with citations:

1. Delete existing `bill_briefs` rows (keeps `bill_documents` intact)
2. Re-queue all documents via a one-off script similar to `queue_docs.py`
3. Worker will regenerate using the new cited prompt at 10/minute
4. ~1,000 briefs ≈ 2 hours

This is **optional** — old string briefs render correctly in the UI with no citation chips.