News Analyzer

News Analyzer is a Rust-based microservice that aggregates financial news, clusters articles into Event Bundles, and provides LLM-powered sentiment analysis for traders.

Overview

News Analyzer operates as part of the FinAI trading system pipeline:

Trader BE → News Analyzer → LLM Wrapper → Trader BE → Trader App

Key Responsibilities:

Fetch news articles from NewsAPI based on asset symbols
Group articles into Event Bundles for multi-source analysis
Send to LLM Wrapper for sentiment analysis
Map LLM responses to structured trader API format

Architecture

API Endpoints

GET /health

Health check endpoint.

Response:

{
  "status": "ok"
}

POST /api/v1/news-analysis

Main analysis endpoint. Accepts asset symbols and returns structured sentiment analysis.

Request:

{
  "request_id": "optional-client-request-id",
  "assets": [
    { "symbol": "AAPL" }
  ],
  "time_range_hours": 24,
  "max_articles": 10,
  "language": "en"
}

Response:

{
  "asset": "AAPL",
  "analysis_timestamp": "2026-05-30T12:00:00Z",
  "overall_sentiment_score": 0.75,
  "confidence_level": 0.82,
  "variance": {
    "status": "Low",
    "type": "Interpretive",
    "note": null
  },
  "key_drivers": [
    {
      "event": "iPhone demand outlook",
      "impact": "High",
      "sentiment": "Bullish",
      "description": "Supplier indicates strong demand"
    }
  ],
  "summary": "Mildly Bullish — Weak Signal",
  "sources": [
    {
      "source": "reuters.com",
      "relevance": 0.9,
      "sentiment": 0.8,
      "url": "https://reuters.com/article/123"
    }
  ],
  "headline_verdict": "Mildly Bullish — Weak Signal. Recent news is slightly positive.",
  "time_horizon": "Medium-term theme",
  "likely_impact": "Medium",
  "risks_and_counterpoints": "No earnings data analysed.",
  "watch_next": ["Apple earnings", "iPhone sales data"]
}

POST /analyze/batch

Batch analysis returning SentimentAnalysisResponse (raw LLM format).

Request:

{
  "asset_name": "BTC",
  "event_bundles": [
    {
      "id": "1",
      "name": "Fed Rate Decision",
      "sources": [
        {
          "label": "Source A",
          "url": "https://reuters.com/btc",
          "title": "Fed holds rates",
          "content": "The Fed kept rates unchanged..."
        }
      ]
    }
  ]
}

POST /analyze/batch/trader

Same as batch but returns TraderApiResponse format.

Data Models

Request DTOs

Model	Fields	Description
`AssetRef`	`symbol: String`	Asset identifier
`NewsAnalysisRequest`	`request_id`, `assets`, `time_range_hours`, `max_articles`, `language`	Main request
`BatchAnalysisRequest`	`asset_name`, `event_bundles`	Internal batch for LLM

Response DTOs

Model	Fields	Description
`TraderApiResponse`	Full trader response	Final response for Trader BE
`SentimentAnalysisResponse`	Raw LLM response	Direct from LLM Wrapper
`VarianceInfo`	`status`, `type`, `note`	Source variance analysis
`KeyDriver`	`event`, `impact`, `sentiment`, `description`	Key market drivers
`SourceInfo`	`source`, `relevance`, `sentiment`, `url`	Source metadata

LLM Integration

Model Priority

gemini-3.5-flash (primary)
gemini-3-flash-preview
gemini-2.5-flash

Response Schema

The LLM is configured with JSON schema validation for structured output:

{
  "asset_name": "string",
  "overall_sentiment": "Bullish" | "Bearish" | "Neutral" | "Mixed",
  "confidence_score": 0.0-1.0,
  "event_analyses": [...],
  "bias_flags": [...],
  "headline_verdict": "string",
  "time_horizon": "Short-term noise" | "Medium-term theme" | "Long-term business signal",
  "likely_impact": "Low" | "Medium" | "High",
  "risks_and_counterpoints": "string",
  "watch_next": ["string"]
}

Retry Logic

On JSON parse failure:

First attempt with full system prompt
Retry with simplified prompt and temperature: 0.1
Fail if retry also fails

Source Authority Tiers

Tier	Weight	Examples
Tier 1	1.0	NYSE, NASDAQ, CME, Bloomberg, Reuters
Tier 2	0.7	WSJ, FT, Barron's
Tier 3	0.4	Secondary outlets
Tier 4	0.2	Social media, forums

Sentiment Scoring

Sentiment	Score
Bullish	1.0
Bearish	-1.0
Neutral	0.0
Mixed	0.0

Variance Detection

Interpretive Variance (Medium severity)

Different interpretations of the same facts
Conflicting analyst opinions
Action: Note, reduce confidence by15%

Factual Conflict (High severity)

Direct contradictions in reported facts
Inconsistent timelines
Action: Flag, reduce confidence by 40-70%

Configuration

Variable	Default	Description
`LLM_ENDPOINT_URL`	`http://llm-wrapper:11435`	LLM Wrapper endpoint
`NEWS_API_KEY`	(required)	NewsAPI.org API key
`RUST_LOG`	`info`	Logging level

Module Structure

src/
├── main.rs                    # Entry point, Axum router
├── config/                    # AppConfig
├── domain/                    # Domain models
├── shared/                    # Error handling, logging
├── middleware/                # Request ID injection
├── clients/                   # External service clients
│   ├── llm_wrapper_client.rs
│   └── news_api_client.rs
└── modules/
    ├── health/               # Health check
    └── news_analysis/        # Main module
        ├── controller.rs    # Route handlers
        ├── service.rs       # Business logic
        ├── dto.rs # Data transfer objects
        ├── prompt.rs       # System prompt template
        └── gemini_schema.rs # JSON schemas

Adding New News Sources

Current State

News Analyzer currently uses NewsAPI.org as its primary news source. The source selection is configured via NEWS_API_KEY.

Planned Architecture

Per the project plan (US-3.2 through US-3.7), future versions will support:

Source Registry - Central source profiles with:
- Domain, enabled flag, discovery methods
- Market/asset type mappings
- Crawl rate limits, JS rendering flags
URL Discovery - Multiple discovery methods:
- RSS feeds (primary)
- Sitemaps (secondary)
- Category pages
- Search API fallback
Keyword Expansion - Asset symbol expansion:
- Primary keywords: AAPL, Apple
- Secondary keywords: Apple stock, Apple earnings
- Negative keywords: filter noise
Article Processing:
- HTML cleaning (remove nav, ads, boilerplate)
- Short article filter (less than 200 chars)
- Relevance filtering (keyword overlap)
- MD5 exact deduplication
Event Bundles - Semantic grouping:
- Embedding generation (all-MiniLM-L6-v2)
- Cosine similarity clustering
- 3-5 authoritative sources per bundle

Implementation Path

To add new sources today:

NewsAPI-based: No code changes needed - just use different search queries
New source type: Implement new client in src/clients/
Full crawling pipeline: Follow US-3.2 → US-3.7 implementation order

LLM Wrapper - LLM integration details
Trader BE - Backend service
Trader FE - Frontend application

Overview​

Architecture​

API Endpoints​

GET /health​

POST /api/v1/news-analysis​

POST /analyze/batch​

POST /analyze/batch/trader​

Data Models​

Request DTOs​

Response DTOs​

LLM Integration​

Model Priority​

Response Schema​

Retry Logic​

Source Authority Tiers​

Sentiment Scoring​

Variance Detection​

Interpretive Variance (Medium severity)​

Factual Conflict (High severity)​

Configuration​

Module Structure​

Adding New News Sources​

Current State​

Planned Architecture​

Implementation Path​

Related Documentation​