extraction prompt

prompts/extract_entities.prompt

@@ -1,34 +1,33 @@
-Input cluster JSON:
-{cluster_json}
+Extract a news signal from the headline AND summary. Never return empty entities if names appear in the text.
+
+Return STRICT JSON with EXACT keys: { topic, entities, sentiment, sentimentScore, keywords }
 
-You MUST extract a news signal from the headline AND summary. Do not leave entities empty when the text mentions obvious names.
-Task:
-1) infer the best top-level topic
-2) extract concise entities from the cluster
-3) assign sentiment from the wording/context
-4) provide short keywords that justify the classification
+FIELDS:
+- topic: one of [crypto, macro, regulation, ai, other]
+- entities: named people, places, orgs, conflicts, and finance/crypto terms
+  (BTC, ETH, ETF, SEC, ECB, Fed, euro, inflation, rates). Canonical forms. 1–5 words each.
+- sentiment: "positive" | "negative" | "neutral"
+- sentimentScore: float –1.0 to 1.0, consistent with sentiment label
+- keywords: 2–4 thematic tags, 1–2 words each. Noun phrases only (e.g. "drone strikes",
+  "nuclear plant"). Not entity names. Not verb phrases. Not headline fragments.
 
-Entity rules (strict):
-- Use short strings (1-5 words).
-- Include all obvious named entities mentioned in headline or summary: people, countries, regions, organizations, ministries, presidents, leaders, wars/conflicts if named.
-- Also include finance/crypto entities when present: BTC, ETH, Bitcoin, Ethereum, ETF, SEC, ECB, Fed, euro, inflation, rates.
-- Prefer canonical entity forms over aliases when obvious (for example, use full organization or place names where helpful).
-- Do NOT return empty entities if any such names/places appear.
+TASKS:
+1. Infer the best topic.
+2. Extract all named entities from headline and summary.
+3. Assign sentiment from tone and wording.
+4. Choose keywords that capture themes, not entities.
 
-Keyword rules (strict):
-- Each keyword MUST be 1-2 words. Never 3+.
-- Keywords are thematic search tags, NOT headline restatements or verb phrases.
-- Good keywords: noun phrases or named concepts (e.g. "drone strikes", "energy infrastructure", "nuclear plant", "oil refinery").
-- Bad keywords: full headline fragments, verb-heavy phrases, or anything over 2 words.
-- Keywords should capture the *themes* of the story, not repeat entity names already in the entities list.
-- Return 2-4 keywords. Fewer is better than bad ones.
+EXAMPLE:
+Input:  { "headline": "ECB raises rates again as eurozone inflation stays elevated",
+          "summary": "The European Central Bank increased its benchmark rate by 25bps,
+                      citing persistent inflation across the eurozone." }
 
-Sentiment rules:
-- positive: clearly encouraging, improving, or supportive tone
-- negative: clearly alarming, worsening, severe, conflict, loss, risk, warning tone
-- neutral: factual, balanced, or mixed
-- sentimentScore must be a number from -1.0 to 1.0 and should reflect the sentiment label.
+Output: { "topic": "macro",
+          "entities": ["ECB", "European Central Bank", "eurozone", "inflation", "rates"],
+          "sentiment": "negative",
+          "sentimentScore": -0.4,
+          "keywords": ["rate hike", "monetary policy"] }
+
+INPUT:
+{cluster_json}
 
-Return STRICT JSON with EXACT keys only:
-{ topic, entities, sentiment, sentimentScore, keywords }
-where topic is one of [crypto, macro, regulation, ai, other].