Inspiration

Rohde and Schwarz, like many other companies, face a big challenge with making use of the data from log files. It requires expert knowledge in the field and several hours. In critical events, there can be a deluge of information and this increases the time it takes to pinpoint the source of error.

What it does

At its heart is an LLM giving the user insights into large log files. It provides summaries as well as

How we built it

Using a combination of Vector Databases, and locally-run fine-tuned Q&A language model (RoBERTa), we extract key information from each line of the log file using a comprehensive Regex analysis, rewrite the log file in natural language, tokenize the lines using our LLM (RoBERTa), and store the tokens in an ultra-fast vector database (ChromaDB).

Challenges we ran into

The LLMs were hard to train and fine-tune on this peculiar kind of data because it is not in natural human language. The concept of summarization is also difficult as log files can span several days from different sources, and hence it depends a lot on the specific context, what a user will call 'a good summary'.

Accomplishments that we're proud of

We have provided a friendly UI. We have accomplished the core required tasks to an MVP degree.

What we learned

It is our first time working on NLP problems, so we learned a lot about the state of the art, the basic libraries, and the difficulties in the problem space.

What's next for Abyss

Better error handling for edge cases. Parsing the LLM outpts better. the outputs vary a lot and need some postprocessing by observation and traditional programmatic rules.