Figure 3
The figure is a three‑section workflow diagram describing construction‑law‑aware question answering. The top section, labeled “(a) Split 387 documents into knowledge chunks”, begins on the left with a vertical stack of rounded rectangles labeled “C L D‑001”, “C L D‑002”, ellipsis dots, and “C L D‑387”, collectively marked “C L K R”. An arrow labeled “Split” points to a box containing items “Knowledge chunk 1”, “Knowledge chunk 2”, ellipsis, and “Knowledge chunk n”. A second arrow labeled “Embed chunks” leads to a box listing “Knowledge vector 1”, “Knowledge vector 2”, ellipsis, and “Knowledge vector n”. A third arrow leads to a magenta box titled “F A I S S‑formatted vector repository” that shows three example high‑dimensional vectors, such as “0.101, negative 0.002, ellipsis, negative 0.400”, “negative 0.003, negative 0.902, ellipsis, negative 0.007”, ellipsis, and “negative 0.803, 0.005, ellipsis, negative 0.243”, under the heading “Store the vectors”. The middle section is titled “(b) Retrieve question‑relevant knowledge chunks”. On the left, a user icon labeled “User” speaks “Questions” in a bubble, and an arrow labeled “Vectorize” points to a box listing “Question vector 1”, “Question vector 2”, up to “Question vector m”. A horizontal arrow labeled “Compare” connects this box to a box describing retrieved knowledge: rows such as “Relevant knowledge chunk 1‑1 to 1‑3”, “Relevant knowledge chunk 2‑1 to 2‑3”, ellipsis, and “Relevant knowledge chunk m‑1 to m‑3”. A side note states, “Retrieve 3 relevant knowledge chunks for each question”, with a loop arrow back up to the F A I S S vector repository, indicating a similarity search between question vectors and stored knowledge vectors. The bottom section is labeled “Combine question and knowledge” on the left and “(c) Input the combined question and retrieved knowledge into G P L L Ms” on the right. Two large ovals represent batched inputs: the top oval shows “Question 1” paired with “Relevant knowledge chunk 1‑1”, “Relevant knowledge chunk 1‑2”, and “Relevant knowledge chunk 1‑3”, while the bottom oval shows “Question m” with “Relevant knowledge chunk m‑1”, “Relevant knowledge chunk m‑2”, and “Relevant knowledge chunk m‑3”. Arrows labeled “Input” point from these ovals to a vertical list titled “General‑purpose large language model (G P L L M)”, which enumerates specific models with icons: “Llama‑2‑70 b”, “Text‑davinci‑003”, “G P T‑3.5 Turbo”, “G P T‑4”, “Chat G L M 2‑6 B”, “E R N I E‑Bot‑turbo”, and “E R N I E‑Bot 4.0”. On the far right, a bubble labeled “Answers” indicates the generated outputs.The process of leveraging CLKR to empower GPLLM for CLQA. Source(s): Authors’ own work