From 266936a4b508ac57ddd9f7b0ff84142bc642552b Mon Sep 17 00:00:00 2001 From: Daniel O'Connell Date: Sat, 30 Sep 2023 18:02:35 +0200 Subject: [PATCH 1/3] Notify when fetching followups --- api/src/stampy_chat/chat.py | 11 +++++++---- api/tests/stampy_chat/test_chat.py | 17 +++++++++-------- web/src/hooks/useSearch.ts | 18 +++++++++--------- web/src/pages/index.tsx | 6 ++++++ 4 files changed, 31 insertions(+), 21 deletions(-) diff --git a/api/src/stampy_chat/chat.py b/api/src/stampy_chat/chat.py index 3a417e6..96e67db 100644 --- a/api/src/stampy_chat/chat.py +++ b/api/src/stampy_chat/chat.py @@ -226,11 +226,14 @@ def talk_to_robot_internal(index, query: str, mode: str, history: Prompt, sessio logger.interaction(session_id, query, response, history, prompt, top_k_blocks) - # yield done state, possibly with followup questions - fin_json = {'state': 'done'} + yield {"state": "loading", "phase": "followups"} + # yield optional followups followups = multisearch_authored([query, response]) - for i, followup in enumerate(followups): - fin_json[f'followup_{i}'] = asdict(followup) + if followups: + yield {'state': 'followups', 'followups': list(map(asdict, followups))} + + # yield done state + fin_json = {'state': 'done'} yield fin_json except Exception as e: diff --git a/api/tests/stampy_chat/test_chat.py b/api/tests/stampy_chat/test_chat.py index 38f0a32..b1b64f5 100644 --- a/api/tests/stampy_chat/test_chat.py +++ b/api/tests/stampy_chat/test_chat.py @@ -268,7 +268,7 @@ def test_talk_to_robot_internal(history, context): with patch('stampy_chat.chat.get_top_k_blocks', return_value=context): with patch('stampy_chat.chat.multisearch_authored', return_value=followups): with patch('openai.ChatCompletion.create', return_value=chunks): - assert list(talk_to_robot_internal("index", "what is this about?", "default", history)) == [ + assert list(talk_to_robot_internal("index", "what is this about?", "default", history, 'session id')) == [ {'phase': 'semantic', 'state': 'loading'}, {'citations': [], 'phase': 'semantic', 'state': 'loading'}, {'phase': 'prompt', 'state': 'loading'}, @@ -277,12 +277,13 @@ def test_talk_to_robot_internal(history, context): {'content': 'response 2', 'state': 'streaming'}, {'content': 'response 3', 'state': 'streaming'}, {'content': 'response 4', 'state': 'streaming'}, - { - 'followup_0': {'pageid': '1', 'score': 0.231, 'text': 'followup 1'}, - 'followup_1': {'pageid': '2', 'score': 0.231, 'text': 'followup 2'}, - 'followup_2': {'pageid': '3', 'score': 0.231, 'text': 'followup 3'}, - 'state': 'done' - }, + {'state': 'loading', 'phase': 'followups'}, + {'state': 'followups', 'followups': [ + {'pageid': '1', 'score': 0.231, 'text': 'followup 1'}, + {'pageid': '2', 'score': 0.231, 'text': 'followup 2'}, + {'pageid': '3', 'score': 0.231, 'text': 'followup 3'}, + ]}, + {'state': 'done'}, ] @@ -297,7 +298,7 @@ def test_talk_to_robot_internal_error(history, context): ] with patch('stampy_chat.chat.get_top_k_blocks', return_value=context): with patch('openai.ChatCompletion.create', return_value=chunks): - assert list(talk_to_robot_internal("index", "what is this about?", "default", history)) == [ + assert list(talk_to_robot_internal("index", "what is this about?", "default", history, 'session id')) == [ {'phase': 'semantic', 'state': 'loading'}, {'citations': [], 'phase': 'semantic', 'state': 'loading'}, {'phase': 'prompt', 'state': 'loading'}, diff --git a/web/src/hooks/useSearch.ts b/web/src/hooks/useSearch.ts index 44e48ee..919d62d 100644 --- a/web/src/hooks/useSearch.ts +++ b/web/src/hooks/useSearch.ts @@ -12,7 +12,7 @@ import type { const MAX_FOLLOWUPS = 4; const DATA_HEADER = "data: " -const EVENT_END_HEADER = "event: close\n" +const EVENT_END_HEADER = "event: close" type HistoryEntry = { role: "error" | "stampy" | "assistant" | "user"; @@ -29,13 +29,13 @@ export async function* iterateData(res: Response) { if (done) return; const chunk = new TextDecoder("utf-8").decode(value); - if (chunk.startsWith(EVENT_END_HEADER)) return; - for (const line of chunk.split("\n")) { // Most times, it seems that a single read() call will be one SSE "message", // but I'll do the proper aggregation spec thing in case that's not always true. - if (line.startsWith(DATA_HEADER)) { + if (line.startsWith(EVENT_END_HEADER)) { + return; + } else if (line.startsWith(DATA_HEADER)) { message += line.slice(DATA_HEADER.length); // Fixes #43 } else if (line !== "") { @@ -83,12 +83,12 @@ export const extractAnswer = async ( setCurrent({ phase: "streaming", ...result }); break; + case "followups": + // add any potential followup questions + followups = data.followups.map((value) => value as Followup); + break; case "done": - // add any potential followup questions - const followups = Object.entries(data) - .filter(([key]) => key.startsWith("followup_")) - .map(([k, value]) => value as Followup); - return { result, followups }; + break; case "error": throw data.error; } diff --git a/web/src/pages/index.tsx b/web/src/pages/index.tsx index 5cb65ae..0a8c79b 100644 --- a/web/src/pages/index.tsx +++ b/web/src/pages/index.tsx @@ -118,6 +118,12 @@ const Home: NextPage = () => { case "streaming": last_entry = ; break; + case "followups": + last_entry = <> + +

Loading: Checking for followups...

+ ; + break; } return ( From 710d7661cdcda27bc5f6a95c3e0a5882f52314b2 Mon Sep 17 00:00:00 2001 From: Daniel O'Connell Date: Sat, 30 Sep 2023 18:03:08 +0200 Subject: [PATCH 2/3] handle duplicate citations --- api/src/stampy_chat/get_blocks.py | 29 ++++++++++++--------- api/tests/stampy_chat/test_get_blocks.py | 32 +++++++++++++++++------- 2 files changed, 40 insertions(+), 21 deletions(-) diff --git a/api/src/stampy_chat/get_blocks.py b/api/src/stampy_chat/get_blocks.py index 91fe738..9b59956 100644 --- a/api/src/stampy_chat/get_blocks.py +++ b/api/src/stampy_chat/get_blocks.py @@ -1,13 +1,12 @@ import dataclasses import datetime -import itertools import numpy as np import openai import regex as re import requests import time from itertools import groupby -from typing import Iterable, List, Tuple +from typing import Iterable, List from stampy_chat.env import PINECONE_NAMESPACE, REMOTE_CHAT_INSTANCE, EMBEDDING_MODEL from stampy_chat import logging @@ -54,12 +53,12 @@ def parse_block(match) -> Block: date = metadata.get('date_published') or metadata.get('date') - if isinstance(date, datetime.date): - date = date.isoformat() - elif isinstance(date, datetime.datetime): + if isinstance(date, datetime.datetime): date = date.date().isoformat() + elif isinstance(date, datetime.date): + date = date.isoformat() elif isinstance(date, (int, float)): - date = datetime.datetime.fromtimestamp(date).isoformat() + date = datetime.datetime.fromtimestamp(date).date().isoformat() authors = metadata.get('authors') if not authors and metadata.get('author'): @@ -82,16 +81,22 @@ def join_blocks(blocks: Iterable[Block]) -> List[Block]: # that the combined block has the minimum index of the blocks combined. def to_tuple(block): - return (block.id, block.title or "", block.authors or [], block.date or "", block.url or "", block.tags or "") + return (block.title or "", block.authors or [], block.date or "", block.url or "", block.tags or "") def merge_texts(blocks): return "\n.....\n".join(sorted(block.text for block in blocks)) - unified_blocks = [ - Block(*key, merge_texts(group)) - for key, group in groupby(blocks, key=to_tuple) - ] - return sorted(unified_blocks, key=to_tuple) + # There are sometimes duplicates in the dataset, but which have different ids, so the id + # is ignored when making sorting the blocks. + def make_block(key, group): + group = list(group) + # Just use the id of the first item - it doesn't matter that much in this case, as the other data points + # will be the same + return Block(group[0].id, *key, merge_texts(group)) + + blocks = sorted(blocks, key=to_tuple) + blocks = [make_block(key, group) for key, group in groupby(blocks, key=to_tuple)] + return blocks # Get the k blocks most semantically similar to the query using Pinecone. diff --git a/api/tests/stampy_chat/test_get_blocks.py b/api/tests/stampy_chat/test_get_blocks.py index 842f60e..4dbc917 100644 --- a/api/tests/stampy_chat/test_get_blocks.py +++ b/api/tests/stampy_chat/test_get_blocks.py @@ -8,22 +8,22 @@ from stampy_chat.get_blocks import Block, get_top_k_blocks, parse_block, join_bl ({}, {}), # Check dates - ({'date_published': '2023-01-02T03:04:05'}, {'date': '2023-01-02T03:04:05'}), + ({'date_published': '2023-01-01T03:04:05'}, {'date': '2023-01-01T03:04:05'}), ( {'date_published': datetime.fromisoformat('2023-01-02T03:04:05')}, - {'date': '2023-01-02T03:04:05'} - ), - ( - {'date_published': datetime.fromisoformat('2023-01-02T03:04:05').date()}, {'date': '2023-01-02'} ), ( - {'date_published': datetime.fromisoformat('2023-01-02T03:04:05').timestamp()}, - {'date': '2023-01-02T03:04:05'} + {'date_published': datetime.fromisoformat('2023-01-03T03:04:05').date()}, + {'date': '2023-01-03'} ), ( - {'date_published': int(datetime.fromisoformat('2023-01-02T03:04:05').timestamp())}, - {'date': '2023-01-02T03:04:05'} + {'date_published': datetime.fromisoformat('2023-01-04T03:04:05').timestamp()}, + {'date': '2023-01-04'} + ), + ( + {'date_published': int(datetime.fromisoformat('2023-01-05T03:04:05').timestamp())}, + {'date': '2023-01-05'} ), # Check authors @@ -84,6 +84,20 @@ def test_parse_block(match_override, block_override): Block('id3', 'title3', ['author3'], 'date3', 'url3', 'tags3', 'text3'), ] ), + ( + [ + Block('id1', 'title1', ['author1'], 'date1', 'url1', 'tags1', 'text1-1'), + Block('id3', 'title3', ['author3'], 'date3', 'url3', 'tags3', 'text3'), + Block('id1', 'title1', ['author1'], 'date1', 'url1', 'tags1', 'text1-2'), + Block('id2', 'title2', ['author2'], 'date2', 'url2', 'tags2', 'text2'), + Block('id1', 'title1', ['author1'], 'date1', 'url1', 'tags1', 'text1-3'), + ], + [ + Block('id1', 'title1', ['author1'], 'date1', 'url1', 'tags1', 'text1-1\n.....\ntext1-2\n.....\ntext1-3'), + Block('id2', 'title2', ['author2'], 'date2', 'url2', 'tags2', 'text2'), + Block('id3', 'title3', ['author3'], 'date3', 'url3', 'tags3', 'text3'), + ] + ), ]) def test_join_blocks(blocks, expected): assert list(join_blocks(blocks)) == expected From 42108e8dfc308532948708d377944fec996bfc72 Mon Sep 17 00:00:00 2001 From: Daniel O'Connell Date: Sat, 30 Sep 2023 18:13:56 +0200 Subject: [PATCH 3/3] Use textareas for previous questions --- web/src/components/entry.tsx | 8 +++----- web/src/pages/index.tsx | 2 +- 2 files changed, 4 insertions(+), 6 deletions(-) diff --git a/web/src/components/entry.tsx b/web/src/components/entry.tsx index d9316b6..10e8eb3 100644 --- a/web/src/components/entry.tsx +++ b/web/src/components/entry.tsx @@ -9,14 +9,12 @@ import { ShowAssistantEntry } from "./assistant"; import { GlossarySpan } from "./glossary"; import Image from "next/image"; import logo from "../logo.svg"; +import TextareaAutosize from 'react-textarea-autosize'; export const User = ({ entry }: { entry: UserEntry }) => { return ( -
  • -

    - {" "} - {entry.content}{" "} -

    +
  • +
    • ); }; diff --git a/web/src/pages/index.tsx b/web/src/pages/index.tsx index 0a8c79b..8e01623 100644 --- a/web/src/pages/index.tsx +++ b/web/src/pages/index.tsx @@ -121,7 +121,7 @@ const Home: NextPage = () => { case "followups": last_entry = <> -

    Loading: Checking for followups...

    +

    Checking for followups...

    ; break; }