What Is LLM Sentiment and How It Is Calculated

LLM: The AI Behind the Scores

LLM stands for Large Language Model — a type of artificial intelligence trained on vast amounts of text that can understand context, nuance, and meaning. SentiSignal uses Google Gemma-3 as its primary LLM, accessed through a LiteLLM gateway for flexible routing and rate management.

The LLM is the primary sentiment engine. It reads each article's full text and produces structured, detailed sentiment analysis — not just a single number, but a complete breakdown of quality, credibility, importance, and per-asset scores.

The Two-Stage Process

Processing every article through a full LLM analysis is expensive (it consumes API tokens and takes 2-5 seconds per article). So SentiSignal uses a two-stage pipeline to be efficient:

Stage 1: Prefilter

The article's title and snippet (up to 400 characters) are sent to the LLM with a simple question: is this article relevant to any tracked asset? The LLM returns a boolean (isRelevant: true/false).

For crypto assets specifically, an algorithmic prefilter runs even before the LLM check. It uses keyword matching — looking for symbol names, aliases, and industry terms — and rejects roughly 40-50% of clearly irrelevant articles before any AI processing occurs. This saves significant cost and processing time.

Stage 2: Deep Analysis

Articles that pass the prefilter are crawled for their full content (up to 6,000 characters of cleaned text). The title and full content are then sent to the LLM with a structured prompt that requests:

• generalSentiment — Overall article sentiment from -1.0 to +1.0
• Per-asset sentiment scores — Individual scores for each relevant symbol mentioned in the article
• quality_score (0–1) — How well-written and substantive the article is
• credibility_score (0–1) — How trustworthy the source and claims appear
• importance_score (0–1) — How likely the article is to impact markets
• reasoning — A text explanation of why the LLM assigned these scores
• ai_description — A 4-5 sentence summary of the article

The Sentiment Scale

The LLM returns sentiment on a continuous scale from -1.0 to +1.0. Here is how to interpret the values:

• -1.0 — Extremely bearish: Panic, crash, regulatory crackdown, existential threats
• -0.5 — Bearish: Significant concerns, setbacks, negative developments
• 0.0 — Neutral: Factual reporting, balanced coverage, no clear directional lean
• +0.5 — Bullish: Partnerships, positive earnings, favorable regulatory signals
• +1.0 — Extremely bullish: Mass adoption, all-time highs, transformative wins

Per-Symbol Nuance

One of the LLM's most powerful capabilities is per-symbol sentiment. A single article can have different sentiment scores for different assets. For example:

• An article titled "Fed Cuts Interest Rates Amid Inflation Concerns" might score +0.6 for GOLD (lower rates support gold prices) but -0.3 for USD (rate cuts weaken the dollar).
• An article about "Ethereum Scaling Breakthrough" might score +0.7 for ETH but +0.2 for BTC (positive for crypto broadly, but the news is ETH-specific).

This per-symbol analysis is what makes SentiSignal's LLM approach fundamentally different from simple keyword-based sentiment tools that assign one score to the entire article.

Where the Data Lives

The LLM analysis results are stored in two database tables:

• news_items_unified.general_sentiment — The overall article sentiment score
• asset_sentiments_unified.sentiment_llm — Per-symbol sentiment scores, one row per asset mentioned in the article

LLM vs VADER

The LLM is the primary engine — slower and more expensive, but far more accurate. It understands sarcasm, context, domain-specific language, and can differentiate between "Bitcoin crashed" (bearish) and "Bitcoin crashed through resistance" (bullish). VADER, by contrast, would rate both as negative due to the word "crashed." For a full comparison, see What Is VADER and How It Differs from LLM.

Why It Matters

Every sentiment score you see on a SentiSignal chart, in an article, brief, or experiment result traces back to this LLM pipeline. Understanding how it works helps you interpret the data more effectively — and recognize situations where the model might struggle (highly ambiguous articles, very short text, or non-English content that was machine-translated).

Next, learn about the complementary engine: What Is VADER and How It Differs from LLM.