View source
Visit project1 Background and motivation
FinSight was developed for Inter IIT Tech Meet 13.0 to answer questions about long financial documents with supporting citations. The difficulty was not simply generating a fluent response. A useful answer had to retrieve the relevant passage, preserve distinctions such as company and reporting year, and let the reader inspect the source. I worked on the pipeline design and evaluation loops for the team’s system.
2 Document processing and retrieval
Documents are parsed, divided into retrievable chunks, and indexed with embeddings and metadata. Retrieval combines similarity with constraints that help keep a question attached to the intended document context. Pathway supports the data flow, while the pipeline preserves references needed to connect an answer with its evidence.
A financial question may require more than one retrieval step. The agentic workflow can decompose it into subquestions, retrieve relevant material for each, and evaluate whether the context is sufficient before generation. This makes the retrieval strategy part of the reasoning process rather than a fixed search performed once at the beginning.
3 Orchestration and evaluation
LangGraph organizes the workflow into nodes and transitions for retrieval, relevance checks, generation, and evaluation. GPT-4 was used for reasoning in the competition system, and language-model evaluators helped inspect answer quality and tune retrieval settings. The goal of those checks was to expose failures early, particularly answers that appeared plausible but were poorly supported.
The repository also contains a product interface with spaces, uploads, document storage, and chat. Its full mode requires configured model and embedding providers; a separate demo mode retains the interface while leaving AI operations unavailable without working keys. This distinction matters when assessing the running demo versus the underlying pipeline.
4 Outcome and limitations
The team placed first in the competition. That outcome records the competition result, while the technical contribution is the document-to-evidence workflow and its evaluation structure. Citations and automated graders help inspection but do not guarantee that every numerical interpretation is correct. Readers still need to verify that a cited passage supports the particular claim made in an answer.