In RAG, attached chunks are joined without separator, potentially leading to concatenations that are misleading for the LLM

[Boscop]

There no separation inserted between the retrieved chunks when they get inserted into the user prompt (considering that the chunks could be split mid-sentence and are usually trimmed of whitespace)?

impl CompletionRequest {
    pub(crate) fn prompt_with_context(&self) -> String {
        if !self.documents.is_empty() {
            format!(
                "<attachments>\n{}</attachments>\n\n{}",
                self.documents
                    .iter()
                    .map(|doc| doc.to_string())
                    .collect::<Vec<_>>()
                    .join(""),
                self.prompt
            )

Without separation, two incomplete sentences (the ending of one chunk and the beginning of another chunk) could form a new sentence that the LLM will interpret, which can give semantically incorrect information!

#184