"Chat with all your emails"
AlphaFinding stuff buried in years of email is painful. Native email search doesn't really help when you need a flight confirmation from six months ago, want to tally up subscription costs, or need that hotel reservation your friend forwarded. Inbox AI lets you upload a Gmail export and ask questions in plain English, like chatting with someone who remembers every message you've ever sent or received.
It parses and indexes your entire email archive, then uses a RAG pipeline to find the most relevant messages for any query. Ask "How much did I spend on flights in 2024?" and it'll dig through thousands of emails, pull out booking confirmations, extract dollar amounts, and give you a clear answer. It goes past keyword matching. It understands context, recognizes entities like merchants, dates, and amounts, and can categorize your emails into groups like travel, finance, shopping, and subscriptions.
It also doubles as an exploration of multi-LLM architectures. You can switch between Claude and GPT-4, making it easy to compare response quality, latency, and cost. Conversations are saved so you can revisit past queries and refine them over time.
Runs the core NLP pipeline: email parsing, text preprocessing, entity extraction, and embedding generation.
The chat interface where you submit queries, view conversations, browse categorized emails, and explore extracted data.
The language understanding layer. Interprets queries, reasons over retrieved email context, and generates answers with citations back to source messages.
Entity extraction, named entity recognition, and topic classification. Pulls dates, amounts, merchant names, and booking details from unstructured email text.
Four-stage pipeline. First, the parsing layer ingests Gmail mbox exports, decodes MIME structures, strips HTML, and normalizes headers into clean document objects. Second, each email gets converted into vector embeddings and stored in a local vector index for fast similarity search. Third, when you submit a query, the RAG retrieval layer encodes the question, runs approximate nearest-neighbor search, and assembles a context window of the most relevant email snippets. Finally, the assembled context plus your question gets sent to the selected LLM (Claude or GPT-4), which generates an answer grounded in the retrieved evidence.
Email is one of the oldest and most inconsistent digital formats still in use. A single Gmail export can contain plain text, nested MIME multipart structures, base64-encoded attachments, HTML-only emails with inline CSS, forwarded chains with mangled headers, and auto-generated messages with wildly varying templates. Getting clean text out of all of these required a lot of edge-case handling: detecting character encodings, stripping boilerplate signatures, reconstructing forwarded threads into coherent timelines.
Pulling structured data like dollar amounts, dates, flight numbers, and merchant names from free-form email text is harder than it sounds. The same info can appear in dozens of formats. A price might be "$1,234.56", "USD 1234.56", or "1,234 dollars" depending on the sender. Dates range from ISO formats to "next Thursday." Reliable extraction required combining regex pattern matching for known templates with NLP-based entity recognition for everything else, plus a validation layer to catch misparses.
Claude and GPT-4 have different pricing, context window sizes, and token-counting behavior, so the same query can cost a lot more on one provider depending on context length. The system uses adaptive context windowing that tailors the number of retrieved email snippets to each provider's token limits and pricing. This keeps answer quality high without surprise cost spikes, while staying transparent about which model is being used and the trade-offs involved.