LLMs are smart, so why don't they know our company's work

LLMs are proficient with public knowledge, but they struggle to answer questions based on fragmented internal information (documents, code, messenger, etc.). Simply improving search performance alone cannot determine whether information is up to date or consistent with the actual code.

To solve this, Toss went beyond search and developed the Topic system, which manages trustworthy context that both LLMs and people can use together. Topic serves as a common layer that manages the source, relationships, recency, and conflict status of information.

The core of Topic is normalizing three sources into a single ContentUnit.

• Documents: Preserves the heading hierarchy of Markdown to maintain context.
• Internal messenger: Summarizes at the Thread level rather than individual messages, preserving the conclusions and decision status of discussions.
• Code: Instead of simple text splitting, it is structured and managed at the symbol and behavior unit level.

Also, rather than simply compressing reliability into a single score, it is verified against six axes, including source, recency, and conflicts. Through this, it clearly identifies what grounds the LLM used to generate an answer, and performs sophisticated quality control that hands off ambiguous decisions to humans.