Cyber-Lenin Improvement Log (260216)
- Simple LangGraph based
- Crawling documents from early 20th century revolutionaries and loading into vector DB (embedding model: BGE-m3)
- Also loading 21st century webzine articles
- If current data is needed, obtains information through web search
- LLM API is constantly being changed
-> OpenAI 4o doesn't have much vitality
-> Dolphin-Mistral-70b's Korean proficiency was too low.
-> Qwen 3.5 Plus performance was decent, but when it encounters politically sensitive parts, it immediately throws a forced block error (link to the article that caused the forced block)
-> Now switched to Gemini-2.5 Flash (strategic planning and final answer generation use 2.5 Flash version, simple utility tasks like identifying question intent use the lightweight model 2.5 Flash-Lite).
- Crawling documents from early 20th century revolutionaries and loading into vector DB (embedding model: BGE-m3)
- Also loading 21st century webzine articles
- If current data is needed, obtains information through web search
- LLM API is constantly being changed
-> OpenAI 4o doesn't have much vitality
-> Dolphin-Mistral-70b's Korean proficiency was too low.
-> Qwen 3.5 Plus performance was decent, but when it encounters politically sensitive parts, it immediately throws a forced block error (link to the article that caused the forced block)
-> Now switched to Gemini-2.5 Flash (strategic planning and final answer generation use 2.5 Flash version, simple utility tasks like identifying question intent use the lightweight model 2.5 Flash-Lite).