OpenClaw Sentiment Assistant Work Report

Work report for the OpenClaw Public Opinion Assistant — an AI-driven, automated sentiment monitoring and analysis system. Report date: March 17, 2026.

I. System Overview

1.1 Positioning

The OpenClaw Sentiment Assistant enables end-to-end operations — news collection, semantic clustering, trend ranking — through natural-language dialogue.

Core value:

• No programming required; conversational operation;
• Automatic deduplication to avoid duplicated-information interference;
• Popularity ranking for quick identification of trending topics.

1.2 Architecture & Data Flow

Data collection → semantic analysis (Zhipu Embed) → vector-based deduplication & clustering → clustering statistics → SQLite storage → popularity ranking → Feishu group push.

II. Core Technology Selection

2.1 Why the Web-Pilot Search Skill?

Criteria	Web-Pilot	Traditional crawler
Anti-scraping	Handled automatically	Manual maintenance
Dynamic content	Supports JS rendering	Static pages only
Maintenance cost	Low (service-level)	High (code changes per page update)
Deployment	Plug-and-play	Environment configuration required

As a search skill inside the OpenClaw ecosystem, Web-Pilot works out of the box and stays stable under anti-scraping measures.

2.2 Why Zhipu Embedding-2?

• Original plan: local BERT model (sentence-transformers);
• Issue: dependency conflicts inside the container (incompatible PyTorch / safetensors versions);
• Re-evaluation: Zhipu Embedding-2 provides an API of equivalent capability;
• Final choice: the API route — lower operating cost, better stability.

Technical specifications: model embedding-2 (BERT series), vector dimension 1024, cosine similarity.

2.3 Why the Leader Clustering Algorithm?

Algorithm	Strengths	Weaknesses	Fit
K-Means	Fast	Needs predefined K; iterative convergence	Known cluster count
DBSCAN	Arbitrary shapes, handles noise	Parameter-sensitive (eps, minPts)	Spatial clustering
Hierarchical	Dendrograms, visualization	O(n³) complexity	Small datasets
Leader (used here)	Simple and fast; order-dependent	—	Real-time clustering, news dedup

Leader algorithm: traverse data sequentially; each unvisited point becomes a “leader”, and points similar to the leader join its cluster.

III. System Features

3.1 Conversational Interface

Users trigger analysis with a natural-language message in a Feishu group:

User: @openclaw Sentiment Assistant, analyze public sentiment. System: automatically performs data collection → clustering → ranking → report generation.

3.2 Output Example

[Today's Top Sentiment Rankings 2026-03-17]
[Ranked by topic popularity (number of related news items)]

No. 1  Crypto trading firm BlockFills files for bankruptcy
       Related news: 2 items | Popularity: 100
No. 2  Institutions showed 'diamond hands' during Bitcoin's 50% pullback
       Related news: 2 items | Popularity: 100

[Statistics] Original news: 27 | Merged topics: 24 | Top topics: 2
Model: Zhipu embedding-2 | Threshold: 0.85

3.3 Core Computation

For each incoming news item, compute cosine similarity against existing items; similarity ≥ 0.85 joins the same topic; remaining unclassified items are re-processed as new leaders. Topics are ranked by the number of member news items (popularity).

IV. Technical Implementation

4.1 Component Specifications

Component	Choice
AI platform	OpenClaw conversational AI control center
Vector model	Zhipu embedding-2
Storage	SQLite
Language	Python 3 standard library — zero dependencies, easy to deploy

4.2 Code Characteristics

• Zero external dependencies: only urllib, json, sqlite3 from the standard library;
• In-house vector math: pure-Python dot product, norm and cosine similarity;
• Container-friendly: no pip install; runs directly.

V. Application Scenarios

• Daily sentiment briefing: scheduled collection, analysis and push;
• Trending event tracking: hottest topics by news count;
• Competitor monitoring: aggregate news by specific keywords.

VI. Next Steps

Phase	Content	Status
MVP	Collection + clustering + ranking; Feishu testing & conversational validation	Completed
In progress	Scheduled tasks (cron automation)	Configuring
Planned	Expand sources to Weibo, Twitter/X, Reddit	In development

VII. Summary

The OpenClaw Sentiment Assistant achieves zero-threshold sentiment monitoring through three core capabilities: conversational AI, semantic clustering and popularity ranking. Replacing local BERT with Zhipu Embedding-2 significantly reduces deployment and maintenance cost.

System documentation for research purposes; not investment advice.