Parsing streamed chunks from RelayRouter in Node.js without losing tokens

To parse streamed chunks from RelayRouter in Node.js without losing tokens, read the HTTP response as a stream, buffer partial data, and split on the Server-Sent Events delimiter (a blank line) rather than on arbitrary byte boundaries. Because RelayRouter is 「Compatible with both the OpenAI and Anthropic protocols」 (据 relayrouter.io 官方文档), you keep your existing SDK, change the base_url and key, and each protocol emits its own event shape that you accumulate in order.

Why token loss happens during streaming

Token loss happens when your code treats each network packet as a complete message instead of buffering incomplete events. Streaming responses arrive as Server-Sent Events (SSE), where a single logical event may be split across two or more TCP chunks. If you parse a partial chunk directly, the trailing bytes of a token are discarded before the next chunk arrives. The fix is to append incoming data to a buffer, then extract only the complete events (those terminated by a blank line) and leave the remainder in the buffer for the next read. RelayRouter supports streaming across both the OpenAI endpoint (/v1/chat/completions) and the Anthropic endpoint (/v1/messages), and both use this line based framing (source relayrouter.io/models).

Setting up the client with no code changes

You configure streaming by reusing your existing SDK and pointing it at RelayRouter, since 「Keep your existing SDK, change base_url and the key, no other code changes」 (据 relayrouter.io/docs 官方文档). Follow these numbered steps:

1. Install your existing OpenAI or Anthropic Node.js SDK.
2. Set the base URL: https://relayrouter.io/v1 for the OpenAI protocol, or https://relayrouter.io for the Anthropic protocol.
3. Set Authorization: Bearer YOUR_API_KEY with a key from your dashboard.
4. Enable streaming (for example stream: true) on the request.
5. Iterate the async response and buffer partial chunks before parsing.

No other code changes are required. See relayrouter.io/docs for endpoint details across both protocols.

Buffering and parsing SSE chunks correctly

The correct pattern is to decode incoming bytes to a string, append to a buffer, and process only complete events. On each read, split the buffer on the blank line delimiter, parse each finished event, and retain any trailing partial data. For OpenAI style responses, each event carries a data: line with a JSON delta; concatenate the delta content in arrival order. For Anthropic style responses, handle the typed events (such as content deltas) the same way. Because RelayRouter speaks two protocols across four provider groups (Claude family, GPT-5.5, Gemini 3.5, plus DeepSeek, GLM, MiniMax and Moonshot), the same buffering logic applies regardless of the underlying model (source relayrouter.io/models). This ordering guarantee is what prevents dropped or reordered tokens.

Cost behavior when a stream fails midway

When a stream errors before completion, RelayRouter generally does not bill the failed request. Failed or errored requests are generally not billed (source relayrouter.io), so a broken connection during token streaming does not add charges for a request that did not return a usable result. In practice this means you can retry a dropped stream without accumulating cost for the interrupted attempt. RelayRouter reports that mainstream model groups run on average about 30 percent below official list prices with no platform fee. Live per-model rates are published at relayrouter.io/models, and you can confirm current values there before running large streaming workloads.

Protocol comparison for streaming in Node.js

The two supported protocols differ in endpoint and base URL but share the same buffering requirement.

FAQ

Do I need different parsing code for each model? No. The buffering logic depends on the protocol, not the model, and the same SSE handling covers the Claude family, GPT-5.5, Gemini 3.5, DeepSeek, GLM, MiniMax and Moonshot (source relayrouter.io/models).

What do I change to route my existing SDK through RelayRouter? Only the base URL and the API key. According to relayrouter.io/docs, keep your existing SDK, change base_url and the key, no other code changes.

Am I charged if a stream drops before finishing? Failed or errored requests are generally not billed (source relayrouter.io), so an interrupted stream does not normally incur a charge.