xLLM by BondingAI: 5–min Demo

⋅

This document features screenshots from a quick live demo intended to potential investors and clients interested in seeing how our production platform works. Contact the author for a live presentation and discussion. Here I run a typical short session on the Nvidia corpus, a repository of public financial documents (PDFs).

Overview

  • Figures 1–3 show the chat mode, common to all LLMs. Ours is specialized for your input corpus(es), and thus requires much less infrastructure (tokens, GPUs) while being optimized for your data with efficient distillation. It also generates synthetic prompts with answers, that the user can modify for training purposes.
  • Figures 4–9 show the structured output and its various components. It is the layer below the final response, where we shine: replicable, auditable, hallucination-free, no-Blackbox algorithms with real-time fine-tune and exhaustive yet concise results with normalized relevancy and trustworthiness scores, along with contextual elements: title, categories, tags, time stamp and so on for each summary box.
  • Figures 10–11 show the general menu before starting a search, a chat, or executing an agent. It illustrates user access and the different channels specific to your organization (marketing, finance, sales, IT, legal, HR and so on).
  • Figure 12 shows specific agents (to be added to production), in this case automated time series and geospatial analyses on retrieved Excel spreadsheets and PDFs, for anomaly detection.

Details

  • Figure 4 features the “Card Information” sub-menu in the exploratory mode (structured output), with the summary boxes called cards, their titles, tags and relevancy scores. Contextual elements may vary depending on the corpus. We will add a trustworthiness score rating the source where the information comes from. Note the tags attached to each card, allowing you to search by tag to find other similar cards.
  • Figure 5 features the “View Details” sub-sub-menu focusing on one card from the previous menu, clicked on by the user. Note the precise reference at the top, to specific chunks in the corpus. The bottom window shows the actual raw content in question, in this case an extract from a PDF. It could also consist of tables (spreadsheet or database extracts) or some images.
  • Figure 6 features the “Explainable” sub-menu in the exploratory mode (structured output), with table display, not specific to a card but to the whole prompt. Each entry in the bottom window matches extracts from the prompt to relevant corpus information not necessarily requested in the prompt, to suggest alternative searchers for your next prompts. Each row shows the association between the entry on the left coming from the prompt, and the corresponding match to the right coming from the corpus. The association is measured using proprietary normalized E-PMI.
  • Figures 7 is same as Figure 6 but with graph instead of table display. You can zoom in and out and get detailed information attached to each circle (E-PMI metric and so on).
  • Figures 8 and 9 feature a prompt that initially failed to return any card with a decent relevancy score. The glitch came from some process in production (now fixed), not from xLLM. I included it to illustrate the fact that we test for exhaustivity in the results, a metric missing in standard benchmarking libraries. When we don’t find relevant material, we don’t make up an answer: we tell you by showing low scores.
  • Figure 12 shows the results from the anomaly detection agent, for fraud detection in a litigation case. The hourly measurements show fraud not detected at the daily level. Dots in the US map are zipcodes: the size represents transaction volume, and the color represents the risk. This case involves an Excel repository with automated extraction, SQL querying, and insights production.

See all the figures in the PDF version of this article, here.

Reference

Liked Liked