AnythingLLM's agent summarizer tool now walks document content one chunk at a time instead of relying on a LangChain map-reduce chain that failed to converge with small context window models. The new approach chunks at 45% of the model's context window, extracts key points per section while carrying forward prior summaries, and asks the user before continuing past the first three sections.
Replaces the LangChain map-reduce summarization chain with a sequential chunk-walk algorithm that limits each chunk to 45% of the model's context window, accumulates key points across chunks, and adds user-confirmation after the first three sections so long documents do not silently block the chat.
Users with small-context-window local models can now summarize long documents without the agent entering an infinite loop, so operations that previously failed silently now complete reliably. The new chunk-walk approach carries prior key points forward for continuity and prompts the user before processing the remainder of a long document, preventing chat lockups. Operators should monitor whether the 45% chunk ratio causes truncation or quality loss on very short context models, and whether the user-confirmation prompt becomes a friction point for automated or batch workflows.