Bring relevant details forward
The app keeps the latest turns and ranks older exchanges by word overlap with the new question. It is a lightweight retrieval heuristic, not a trained attention model.
A chat should feel like one conversation, not a series of fresh starts. Baatwise is an early-stage project exploring how context can make the next reply more relevant.
People add details as they go: a constraint, a preference, a correction. We’re exploring how a conversational assistant can bring those signals into the next answer.
The app keeps the latest turns and ranks older exchanges by word overlap with the new question. It is a lightweight retrieval heuristic, not a trained attention model.
Set tone, answer detail, and English, Hindi, or Telugu. Claude Sonnet 5.5 is the default; Claude Haiku 4.5 is available as a lower-cost option.
Attach a PDF up to 10 MB and 50 pages. The app extracts at most 18,000 characters of text; scanned PDFs need OCR first.
The repository now contains a runnable Streamlit chat powered by Anthropic’s Claude API. It combines recent messages with a small number of earlier turns selected for lexical relevance, so a follow-up can use useful details without replaying the entire transcript.
Use Claude Sonnet 5.5 for balanced replies or choose Haiku 4.5 as the lower-cost option. You supply your own Anthropic API key.
Recent exchanges stay available while older turns are selected by word overlap with the current question. There are no accounts or saved profiles.
Web search is off by default. PDF text can be added for a session. Search, model tokens, and your API plan may incur charges.
EARLY PROTOTYPE · SESSION-ONLY CHAT · NOT A PRODUCTION SERVICE
Prompts and extracted PDF text are sent to Anthropic. Baatwise does not save them to its own database. Web search may add separate charges. No customer, reliability, or performance claims have been validated.
Baatwise brings an everyday Indian word into a simple idea: a useful AI chat should keep track of what the conversation is about. The visual identity follows the same thought—one continuous line, with room for the next turn.
Read the setup guide, run the app with your own Anthropic API key, and follow the open-source experiment.
Feedback: raghu@baatwise.in.net · Founder on LinkedIn ↗