Duriin-API/README.md

# duriin_api

Node.js Fastify server that ingests news articles from RSS, Google News RSS, SEC EDGAR 8-K filings, Alpha Vantage News Sentiment, Finnhub company news, and GDELT into a local SQLite archive.

## Setup

1. Install dependencies:
   ```bash
   npm install
   ```
2. Edit `config.json` with your API keys, tickers, RSS feeds, Google News settings, and schedules.
3. Start the server:
   ```bash
   npm start
   ```

The server listens on the host and port defined in `config.json`.

## How the data pipeline works

On startup the server:

1. Opens the SQLite database.
2. Registers the article and status routes.
3. Starts the HTTP server.
4. Immediately runs all ingestion sources once.
5. Starts the cron scheduler for recurring ingestions, content backfill, and embedding backfill.

When a new article is inserted:

- the record is written immediately with `title`, `description`, `url`, `source`, and timestamps
- `content` and `image` start as `null`
- full article extraction runs asynchronously after insert
- vector embeddings are generated later, after title, description, and content are all available

## API overview

All exposed endpoints are `GET` endpoints.

### `GET /`

Simple health check.

**Response**
```json
{ "ok": true }
```

Use this to confirm the server is running, not to inspect ingestion state.

### `GET /articles`

Returns articles from the `articles` table. Only articles that are considered **usable** are exposed — meaning they have non-empty `content`, a stored embedding, and are not index/category pages. Behavior changes based on the query params you send.

#### Query params

##### `keyword`

Plain keyword search.

- matches `title`, `description`, and `content`
- uses SQL `LIKE`
- works like substring matching, not semantic search
- best when you want literal words or phrases to appear in the article text

Example:
```http
GET /articles?keyword=earnings
```

##### `source`

Exact match on the stored `source` field.

Example:
```http
GET /articles?source=rss
```

##### `from`

Only returns rows where `pub_date >= from`.

Example:
```http
GET /articles?from=2025-01-01T00:00:00.000Z
```

##### `to`

Only returns rows where `pub_date <= to`.

Example:
```http
GET /articles?to=2025-01-31T23:59:59.999Z
```

##### `limit`

Number of rows to return.

- default: `20`
- max: `100`

Example:
```http
GET /articles?limit=10
```

##### `offset`

Pagination offset.

- default: `0`

Example:
```http
GET /articles?limit=10&offset=20
```

##### `similar_to_article`

Runs vector similarity search instead of normal list mode.

- value must be an existing article ID
- the server looks up that article's embedding
- nearest-neighbor search runs in `sqlite-vec`
- the source article is excluded from the result set
- each result includes a `distance` field
- lower `distance` means more similar
- returns `404` if the article has no stored embedding

Example:
```http
GET /articles?similar_to_article=123&limit=5
```

Not found response:
```json
{ "error": "Embedding not found for article" }
```

##### `semantic`

Semantic search by meaning, not exact wording.

- use this when you want conceptually related results
- unlike `keyword`, the words do not need to appear literally in the article text
- the query text is normalized before embedding
- query embeddings are cached in SQLite
- on cache miss, the server requests an embedding from OpenRouter
- nearest article matches are returned from the embedding index
- each result includes a `distance` field
- lower `distance` means a closer semantic match
- returns `400` if `semantic` is empty

Example:
```http
GET /articles?semantic=ai chip demand&limit=10
```

Bad request response:
```json
{ "error": "Semantic query must not be empty" }
```

##### `include_embedding`

Explicitly rejected on `/articles`.

Response:
```json
{ "error": "Embeddings are not returned directly. Use similar_to_article for vector search." }
```

#### General behavior

- If `semantic` is present, semantic search is used.
- Else if `similar_to_article` is present, similarity search is used.
- Otherwise normal list/search mode is used.
- `keyword` is literal keyword matching.
- `semantic` is semantic matching by meaning.
- Normal list/search results are ordered by `COALESCE(pub_date, ingested_at) DESC, id DESC`.
- `from` and `to` are compared against stored publication timestamps, so ISO-8601 values are the safest input.
- `source` must match the stored source name exactly.
- `keyword` is substring matching, not full-text search.

#### Normal list/search response shape

```json
[
  {
    "id": 123,
    "title": "...",
    "description": "...",
    "content": "...",
    "image": "...",
    "url": "...",
    "normalized_title": "...",
    "source": "rss",
    "pub_date": "2025-01-01T12:34:56.000Z",
    "ingested_at": "2025-01-01T12:35:10.000Z"
  }
]
```

#### Similarity/topic search response shape

```json
[
  {
    "id": 456,
    "title": "...",
    "description": "...",
    "content": "...",
    "image": "...",
    "url": "...",
    "normalized_title": "...",
    "source": "rss",
    "pub_date": "2025-01-02T09:00:00.000Z",
    "ingested_at": "2025-01-02T09:00:10.000Z",
    "distance": 0.1234
  }
]
```

#### Combined example

```http
GET /articles?keyword=earnings&source=rss&from=2025-01-01T00:00:00.000Z&limit=10&offset=0
```

### `GET /articles/:id`

Returns one article by numeric ID.

**Behavior**

- Looks up the article directly in SQLite.
- Same usability filter as the list endpoint — returns `404` if the article exists but is not usable.
- Returns the same article fields as normal `/articles` list mode.
- Does not return embedding data.
- Returns `404` if the ID does not exist.

**Example**
```http
GET /articles/123
```

**Not found response**
```json
{ "error": "Article not found" }
```

### `GET /status`

Returns ingestion and archive summary information.

**Response fields**

- `total`: total number of rows in `articles` across all sources
- `usable`: articles that have content, an embedding, and are not index pages
- `lastIngestionBySource`: in-memory timestamps of the last successful batch run per source
- `bySource`: per-source breakdown, each with `total` and `usable` counts

**Important detail**

`lastIngestionBySource` is kept in memory, so it resets when the process restarts.

**Example response**
```json
{
  "total": 10234,
  "usable": 8700,
  "lastIngestionBySource": {
    "rss": "2025-01-02T10:00:00.000Z",
    "gdelt": "2025-01-02T10:05:00.000Z"
  },
  "bySource": {
    "alphavantage": { "total": 120, "usable": 98 },
    "edgar": { "total": 88, "usable": 70 },
    "finnhub": { "total": 400, "usable": 360 },
    "gdelt": { "total": 2100, "usable": 1800 },
    "rss": { "total": 7526, "usable": 6372 }
  }
}
```

## Article field notes

- `image` stores the extracted main image as ultra-compressed base64 WebP.
- `normalized_title` is stored for matching and indexing.
- `source` may be a shared source like `rss`, `googlenews`, `gdelt`, `edgar`, `alphavantage`, or `finnhub`.
- `pub_date` is normalized to ISO-8601 when it can be parsed.
- `ingested_at` is the insert timestamp set by the server.

## Notes

- SQLite archive file defaults to `./archive.sqlite`.
- Deduplication is enforced on `url`; normalized titles are stored and indexed for matching but are not unique.
- `googleNews` accepts `queries`, `topics`, `language`, and `country`, and resolves Google redirect URLs to publisher URLs before ingestion.
- Article body extraction runs asynchronously after insertion, with scheduled retries for rows still missing content.
- Embeddings are generated asynchronously with OpenRouter `perplexity/pplx-embed-v1-0.6b` and indexed in `sqlite-vec` for similarity search.
- Topic search caches normalized query embeddings in SQLite and falls back to OpenRouter on cache miss.
- SEC requests use the configured `User-Agent`.
- Duplicate URLs are skipped rather than inserted again.