(Anthropic LLMs) First two characters of a response are lost

[jwr]

I'm trying out gptel with Anthropic Claude and it seems that with many (but curiously not all, I think?) responses the first two characters are lost somewhere.

For example, with an empty prompt in a conversation:

### 

 seems like you haven't provided any additional context or code. The previous two functions we discussed were:

Or when working in another buffer:

How do I update data in an atom in Clojure?

 Clojure, you can update the of an atom using the `swap! or `reset!` functions.. **Using `swap!`**:
The `swap!` function takes an atom and a function as arguments. It applies the function to the current value of the atom, and sets the atom's value to the result of the function. This is useful when you want to perform a transformation on current value of the atom.
[...]

Note how in the first example the word "It" is missing, and in the second one "In" is missing.

I'm using gptel 0d6264f in "GNU Emacs 29.1 (build 1, aarch64-apple-darwin21.6.0, Carbon Version 165 AppKit 2113.6) of 2023-08-10" (emacs-mac).

[karthink]

1. Are you using gptel-mode/a dedicated chat buffer?
2. Is this an issue only with the Claude model?
3. If possible, could you try the previous commit? (If you installed gptel using straight/elpaca, this should be easy to do. Otherwise don't worry about it)

[jwr]

1. This happens with output to normal buffers as well as in a dedicated gptel chat buffer.
2. I cannot reproduce this problem with eb088f2 and gpt-4-turbo-preview from OpenAI.
3. I went back to eb088f2 and initially responses in buffers were not truncated, but then I tried the dedicated gptel chat buffer and got this:

### Are you familiar with Clojure?

, I'm familiar with Clojure. It's a modern Lisp dialect that runs on the Java Virtual Machine (JVM) and JavaScript engines. Clojure emphasizes functional programming, immutable data structures, and concurrency support through software transactional memory. It has a rich set of data structures and a focus on simplicity and consistency.

### 

But this time it seems that 3 characters ("Yes") are missing?

This is fairly consistent:

### Can you write Clojure code?

, I can write Clojure code. Here's a simple example that defines a function to calculate the factorial of a number:

[jwr]

I was wrong, it's not two characters in front, but some characters in multiple places in the response.

I am looking at a gptel chat buffer where there are several places where 1-5 characters are missing (streaming chunk boundaries?), not just at the beginning, but in the middle of the response as well. I see this with Anthropic only, not with OpenAI or ollama.

(I am back to current head, or 0d6264f)

[karthink]

I am looking at a gptel chat buffer where there are several places where 1-5 characters are missing (streaming chunk boundaries?), not just at the beginning, but in the middle of the response as well. I see this with Anthropic only, not with OpenAI or ollama.

Thank you for the thorough testing, this is very helpful. It's probably a bug in the parser for the Anthropic-API responses. Could you do one more thing to help me track it down?

1. Run (setq gptel-log-level 'info)
2. Use Claude, generate some responses with missing chunks.
3. Paste the contents of the *gptel-log* buffer here. (Check to ensure that the buffer does not contain your API key. At the info log level it shouldn't, but please check anyway)
4. Paste the chat here as well so I can compare the log and the responses as printed in your working buffer.

[jwr]

Unfortunately, I currently can't, as Anthropic locked me out when my $5 credit ran out, and even though I recharged the account and it shows a significant balance in the panel, their API responds with 400 codes. And their support mentions that they respond after 5 days 🤣

If somebody else doesn't help in the meantime, this will have to wait until they tie their shoelaces and get their act together, then I'll be able to get back on it!

[solodov]

I have the same problem, here's my data.

ChatGPT buffer:

### is this on?

, I'm here and ready to assist you. How can I help?

### 

gptel-log

{
  "gptel": "request body",
  "timestamp": "2024-03-07 17:47:00"
}
{
  "model": "claude-3-opus-20240229",
  "messages": [
    {
      "role": "user",
      "content": "is this on?"
    }
  ],
  "system": "You are a large language model living in Emacs and a helpful assistant. Respond concisely.",
  "stream": true,
  "max_tokens": 1024,
  "temperature": 1.0
}
{
  "gptel": "response body",
  "timestamp": "2024-03-07 17:47:05"
}
event: message_start
data: {"type":"message_start","message":{"id":"msg_01CXFfSUxbDeyrqJkwPj1UnU","type":"message","role":"assistant","content":[],"model":"claude-3-opus-20240229","stop_reason":null,"stop_sequence":null,"usage":{"input_tokens":33,"output_tokens":1}}}

event: content_block_start
data: {"type":"content_block_start","index":0,"content_block":{"type":"text","text":""}}

event: ping
data: {"type": "ping"}

event: content_block_delta
data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":"Yes"}}

event: content_block_delta
data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":","}}

event: content_block_delta
data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":" I"}}

event: content_block_delta
data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":"'m"}}

event: content_block_delta
data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":" here"}}

event: content_block_delta
data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":" and"}}

event: content_block_delta
data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":" ready"}}

event: content_block_delta
data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":" to"}}

event: content_block_delta
data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":" assist"}}

event: content_block_delta
data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":" you"}}

event: content_block_delta
data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":"."}}

event: content_block_delta
data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":" How"}}

event: content_block_delta
data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":" can"}}

event: content_block_delta
data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":" I"}}

event: content_block_delta
data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":" help"}}

event: content_block_delta
data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":"?"}}

event: content_block_stop
data: {"type":"content_block_stop","index":0}

event: message_delta
data: {"type":"message_delta","delta":{"stop_reason":"end_turn","stop_sequence":null},"usage":{"output_tokens":19}}

event: message_stop
data: {"type":"message_stop"}

The log does show that the very first event is dropped in the chat.

[karthink]

@solodov Thank you.

It's strange, I've tried all the prompts suggested in this thread so far and I'm not able to reproduce the missing chunk problem. It works fine on the test case. I'm trying to guess the cause from staring at the parser code now.

[karthink]

Since I'm not sure what's causing the parsing problem, I've attempted a fix based on my best guess. Please let me know if it makes a difference.

[solodov]

Looks like my case now works correctly, thanks!

…

On Fri, Mar 8, 2024 at 2:06 AM karthink ***@***.***> wrote: Since I'm not sure what's causing the parsing problem, I've attempted a fix based on my best guess. Please let me know if it makes a difference. — Reply to this email directly, view it on GitHub <#233 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AABN2WRJ2GMQQTMAP35A2LLYXEFKRAVCNFSM6AAAAABEJW6N52VHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMYTSOBUHA3DQNRXGA> . You are receiving this because you were mentioned.Message ID: ***@***.***>

[karthink]

Okay. I'll wait until @jwr can access Claude again and check if the bug still persists before closing this issue.

[jwr]

Sorry, it took a while for Anthropic to figure out that I do have a positive balance in my account after all.

I can now confirm that the bug is gone, and I get full responses from Anthropic models (tested with fbb0ee2).

Thank you!