Remove unused summary/conclusion logic (#1179)

* remove unused summary/conclusion logic

* Set timestamp when concluding answer

* Restore line number offset logic in transcoder, fix tests

---------

Co-authored-by: calyptobai <calyptobai@gmail.com>

server/bleep/src/agent.rs

@@ -91,7 +91,7 @@ impl Drop for Agent {
                             .with_payload("message", "request panicked"),
                     );
                 } else {
-                    self.last_exchange_mut().apply_update(Update::Cancel);
+                    self.last_exchange_mut().apply_update(Update::SetTimestamp);
 
                     self.track_query(
                         EventData::output_stage("cancelled")
@@ -319,8 +319,8 @@ impl Agent {
 
                 let answer = match e.answer() {
                     // NB: We intentionally discard the summary as it is redundant.
-                    Some((answer, _conclusion)) => {
-                        let encoded = transcoder::encode_summarized(answer, None, "gpt-3.5-turbo")?;
+                    Some(answer) => {
+                        let encoded = transcoder::encode_summarized(answer, "gpt-3.5-turbo")?;
                         Some(llm_gateway::api::Message::function_return("none", &encoded))
                     }
 

server/bleep/src/agent/exchange.rs

@@ -2,7 +2,6 @@ use crate::query::parser::SemanticQuery;
 use std::fmt;
 
 use chrono::prelude::{DateTime, Utc};
-use rand::seq::SliceRandom;
 
 /// A continually updated conversation exchange.
 ///
@@ -59,25 +58,10 @@ impl Exchange {
             Update::Article(full_text) => {
                 *self.answer.get_or_insert_with(String::new) = full_text;
             }
-            Update::Conclude(conclusion) => {
-                self.response_timestamp = Some(Utc::now());
-                self.conclusion = Some(conclusion);
-            }
             Update::Focus(chunk) => {
                 self.focused_chunk = Some(chunk);
             }
-            Update::Cancel => {
-                let conclusion = [
-                    "The article wasn't completed. See what's available",
-                    "Your article stopped before completion. Check out the available content",
-                    "The content stopped generating early. Review the initial response",
-                ]
-                .choose(&mut rand::thread_rng())
-                .copied()
-                .unwrap()
-                .to_owned();
-
-                self.conclusion = Some(conclusion);
+            Update::SetTimestamp => {
                 self.response_timestamp = Some(Utc::now());
             }
         }
@@ -88,14 +72,11 @@ impl Exchange {
         self.query.target().map(|q| q.to_string())
     }
 
-    /// Get the answer and conclusion associated with this exchange, if a conclusion has been made.
+    /// Get the answer associated with this exchange.
     ///
-    /// This returns a tuple of `(full_text, conclusion)`.
-    pub fn answer(&self) -> Option<(&str, &str)> {
-        match (&self.answer, &self.conclusion) {
-            (Some(answer), Some(conclusion)) => Some((answer.as_str(), conclusion.as_str())),
-            _ => None,
-        }
+    /// This returns a tuple of `full_text`.
+    pub fn answer(&self) -> Option<&str> {
+        return self.answer.as_deref();
     }
 
     /// Return a copy of this exchange, with all function call responses redacted.
@@ -203,7 +184,6 @@ pub enum Update {
     StartStep(SearchStep),
     ReplaceStep(SearchStep),
     Article(String),
-    Conclude(String),
     Focus(FocusedChunk),
-    Cancel,
+    SetTimestamp,
 }

server/bleep/src/agent/tools/answer.rs

@@ -2,7 +2,6 @@ use std::{collections::HashMap, mem, ops::Range, pin::pin};
 
 use anyhow::{anyhow, Context, Result};
 use futures::StreamExt;
-use rand::{rngs::OsRng, seq::SliceRandom};
 use tracing::{debug, info, instrument, trace};
 
 use crate::{
@@ -77,32 +76,15 @@ impl Agent {
             let fragment = fragment?;
             response += &fragment;
 
-            let (article, summary) = transcoder::decode(&response);
+            let article = transcoder::decode(&response);
             self.update(Update::Article(article)).await?;
-
-            if let Some(summary) = summary {
-                self.update(Update::Conclude(summary)).await?;
-            }
         }
 
-        // We re-decode one final time to catch cases where `summary` is `None`, and to log the
-        // output as a trace.
-        let (article, summary) = transcoder::decode(&response);
-        let summary = summary.unwrap_or_else(|| {
-            [
-                "I hope that was useful, can I help with anything else?",
-                "Is there anything else I can help you with?",
-                "Can I help you with anything else?",
-            ]
-            .choose(&mut OsRng)
-            .copied()
-            .unwrap()
-            .to_owned()
-        });
-
-        trace!(%article, "generated answer");
+        if let Some(article) = self.last_exchange().answer() {
+            trace!(%article, "generated answer");
+        }
 
-        self.update(Update::Conclude(summary)).await?;
+        self.update(Update::SetTimestamp).await?;
 
         self.track_query(
             EventData::output_stage("answer_article")
@@ -223,10 +205,8 @@ impl Agent {
                     content: q,
                 });
 
-                let conclusion = e.answer().map(|(answer, conclusion)| {
-                    let encoded =
-                        transcoder::encode_summarized(answer, Some(conclusion), "gpt-4-0613")
-                            .unwrap();
+                let conclusion = e.answer().map(|answer| {
+                    let encoded = transcoder::encode_summarized(answer, "gpt-4-0613").unwrap();
 
                     llm_gateway::api::Message::PlainText {
                         role: "assistant".to_owned(),