feat(dedup): implement two-stage duplicate detection and content rewriter

2026-08-27 19:53:34 +03:30
parent 4e6b3eca6f
commit d823a3749a
4 changed files with 286 additions and 0 deletions
@@ -0,0 +1,138 @@
+import logging
+from typing import List, Optional, Dict, Any
+from db.models import Post, TargetChannel
+from db.repository import Repository
+from core.llm import LLMClient
+from core.dedup import compute_content_hash
+from core.metrics import DUPLICATES_DETECTED_TOTAL
+
+logger = logging.getLogger(__name__)
+
+TAG_EXTRACTION_SYSTEM_PROMPT = """
+You are an AI news analyst and classifier.
+Given a social media/channel post, extract:
+1. "subject": A brief, specific headline/subject (3-8 words).
+2. "tags": A JSON array of 3 to 6 lowercase keywords/topics/entities (e.g. ["ai", "nvidia", "gpus", "hardware"]).
+Respond ONLY in JSON format:
+{
+  "subject": "...",
+  "tags": ["tag1", "tag2", "tag3"]
+}
+"""
+
+DUPLICATE_CHECK_SYSTEM_PROMPT = """
+You are an expert news editor checking for duplicate news stories.
+Given a NEW POST and a list of PREVIOUS POSTS, determine if the NEW POST is covering the same exact event, news item, or story as any of the previous posts.
+
+Respond ONLY in JSON format:
+{
+  "is_duplicate": true/false,
+  "duplicate_of_id": <id of matched previous post or null>,
+  "similarity_reason": "<short explanation of why it is or is not a duplicate>"
+}
+"""
+
+POST_REWRITE_SYSTEM_PROMPT = """
+You are an expert Telegram content creator and copywriter.
+Rewrite the provided post to make it engaging, well-formatted, professional, and clear.
+Use appropriate emojis, clear paragraphs, and markdown formatting.
+Remove any original promotional links, author credits, or watermarks.
+
+Respond ONLY in JSON format:
+{
+  "ai_text": "...",
+  "suggested_target_id": <optional id of best matching target channel or null>
+}
+"""
+
+class AIProcessor:
+    def __init__(self, repo: Repository, llm: Optional[LLMClient] = None):
+        self.repo = repo
+        self.llm = llm or LLMClient()
+
+    async def process_post(self, post_id: int) -> Optional[Post]:
+        post = await self.repo.get_post_by_id(post_id)
+        if not post or not post.raw_text:
+            return post
+
+        raw_text = post.raw_text
+        is_dup = False
+        dup_of_id = None
+        sim_reason = None
+
+        # 1. Exact hash duplicate check
+        content_hash = post.content_hash or compute_content_hash(raw_text)
+        if content_hash:
+            exact_dup = await self.repo.find_duplicate_post_by_hash(content_hash)
+            if exact_dup and exact_dup.id != post.id:
+                is_dup = True
+                dup_of_id = exact_dup.id
+                sim_reason = "Exact match on normalized text/media hash"
+                DUPLICATES_DETECTED_TOTAL.labels(method="hash").inc()
+
+        # 2. Extract Tags and Subject via AI
+        tags = []
+        subject = "General News"
+        try:
+            tag_res = await self.llm.generate_json(
+                prompt=f"Post content:\n\n{raw_text}",
+                system_prompt=TAG_EXTRACTION_SYSTEM_PROMPT,
+                action_name="extract_tags"
+            )
+            subject = tag_res.get("subject", subject)
+            tags = [t.lower().strip() for t in tag_res.get("tags", []) if isinstance(t, str)]
+            await self.repo.update_post_tags(post.id, tags, subject)
+        except Exception as e:
+            logger.error(f"Tag extraction failed for post {post.id}: {e}")
+
+        # 3. Candidate search & Semantic AI Deduplication check (if not already exact dup)
+        if not is_dup and tags:
+            candidates = await self.repo.find_candidate_posts_by_tags(tags, exclude_post_id=post.id, hours_lookback=72, limit=5)
+            if candidates:
+                cand_texts = "\n---\n".join([f"ID {c.id} (Subject: {c.subject}):\n{c.raw_text}" for c in candidates if c.raw_text])
+                prompt = f"NEW POST:\n{raw_text}\n\nPREVIOUS CANDIDATE POSTS:\n{cand_texts}"
+                try:
+                    dup_res = await self.llm.generate_json(
+                        prompt=prompt,
+                        system_prompt=DUPLICATE_CHECK_SYSTEM_PROMPT,
+                        action_name="check_duplicate"
+                    )
+                    if dup_res.get("is_duplicate"):
+                        is_dup = True
+                        dup_of_id = dup_res.get("duplicate_of_id")
+                        sim_reason = dup_res.get("similarity_reason", "AI detected duplicate news topic")
+                        DUPLICATES_DETECTED_TOTAL.labels(method="ai_semantic").inc()
+                except Exception as e:
+                    logger.error(f"Semantic duplicate check failed for post {post.id}: {e}")
+
+        # 4. Rewrite post for our channels
+        ai_text = raw_text
+        suggested_target_id = None
+        targets = await self.repo.get_active_targets()
+        target_info = "\n".join([f"Target ID {t.id}: {t.title} (@{t.username or 'none'})" for t in targets])
+        rewrite_prompt = f"TARGET CHANNELS AVAILABLE:\n{target_info or 'None'}\n\nORIGINAL POST:\n{raw_text}"
+
+        try:
+            rewrite_res = await self.llm.generate_json(
+                prompt=rewrite_prompt,
+                system_prompt=POST_REWRITE_SYSTEM_PROMPT,
+                action_name="rewrite_post"
+            )
+            ai_text = rewrite_res.get("ai_text", raw_text)
+            suggested_target_id = rewrite_res.get("suggested_target_id")
+        except Exception as e:
+            logger.error(f"Post rewrite failed for post {post.id}: {e}")
+
+        # 5. Save AI results into database
+        await self.repo.update_ai_result(
+            post_id=post.id,
+            subject=subject,
+            ai_text=ai_text,
+            tags=tags,
+            suggested_target_id=suggested_target_id,
+            is_duplicate=is_dup,
+            duplicate_of_id=dup_of_id,
+            similarity_reason=sim_reason,
+        )
+
+        return await self.repo.get_post_by_id(post.id)