GPT-4 Tutorial: How to Chat With Multiple PDF Files (~1000 pages of Tesla's 10-K Annual Reports)

Answering questions across multiple massive PDF files—like several years of Tesla 10-K annual reports—becomes practical when each document is converted into searchable embeddings and the system can dynamically route a user’s question to the right year-specific “namespace” in a vector database. The core payoff: a single chat can retrieve relevant passages from 2020, 2021, and 2022 reports, then compute or summarize across them while citing the underlying page sources.

The workflow starts with ingestion. Each PDF is converted from binary into text, then split into chunks because models only handle limited context at a time. Those chunks are embedded into numeric vectors (representations of meaning) and stored in a vector store so retrieval can happen fast. The design uses separate namespaces per year—e.g., Tesla 2020, Tesla 2021, Tesla 2022—so the system can search within the correct subset of documents. Pinecone is used as the vector store, and the embeddings dimension must match the embedding model output (the transcript notes 1536 for OpenAI embeddings). During ingestion, metadata such as page numbers and source references are attached to chunks so answers can be grounded with citations.

Query-time logic is where multi-year analysis is made to work. For a single-year question, the system can search one namespace directly. But cross-year questions require a routing strategy: the system must determine which years the user is referring to, then search multiple namespaces and combine the retrieved context. To avoid hard-coding, GPT-4 is used to extract the years from the user’s question (via a function that returns the referenced years). Those years are mapped to the corresponding namespaces. Once the relevant namespaces are identified, the system retrieves top-matching chunks from each, merges the context, and sends it to GPT-4 to produce a synthesized answer.

The transcript demonstrates this with finance-style prompts such as “What were the risk factors for Tesla 2022?” (with page-level references) and then escalates to questions spanning multiple years, like gross margin changes over three years and revenue growth calculations. In the multi-year case, the system not only pulls supporting excerpts from each annual report but also performs arithmetic based on the retrieved statements (for example, computing revenue growth using consolidated statement figures). The demo also references a front-end and API adaptation, with LangChain chains orchestrating the steps: extracting years, retrieving from Pinecone namespaces, and generating the final response.

Implementation details reinforce operational realism: ingestion is done via scripts that load PDFs from a reports folder, chunk them (the transcript mentions chunk sizes and overlap), and insert vectors in batches because Pinecone has limits (the transcript notes a recommendation around 100 vectors per insert). The result is an architecture that can chat over roughly a thousand pages across multiple PDFs while keeping answers tied to the original documents—turning tedious reading into targeted, source-cited analysis across time.