Skip to content

web search done right

Alice; show claim confidence level

Mikkel Denker18f7ef184271
diff --git a/configs/frontend.toml b/configs/frontend.toml
index ca40cdf7..0930fcc0 100644
--- a/configs/frontend.toml
+++ b/configs/frontend.toml
@@ -5,6 +5,7 @@ prometheus_host = "0.0.0.0:3001"
gossip_addr = "0.0.0.0:3005"
gossip_seed_nodes = ["0.0.0.0:3006", "0.0.0.0:3007"]
crossencoder_model_path = "data/cross_encoder"
+fact_check_model_path = "data/fact_model"
# lambda_model_path = "data/lambdamart.txt"
# qa_model_path = "data/qa_model"
bangs_path = "data/bangs.json"
diff --git a/core/src/alice/mod.rs b/core/src/alice/mod.rs
index 78046c6c..3f38a12e 100644
--- a/core/src/alice/mod.rs
+++ b/core/src/alice/mod.rs
@@ -287,7 +287,7 @@ impl Alice {
}
let mut summarizer = ExtractiveSummarizer::open(summarizer_path, 1)?;
- summarizer.set_window_size(30);
+ summarizer.set_window_size(50);
let summarizer = Arc::new(summarizer);
diff --git a/core/src/api/fact_check.rs b/core/src/api/fact_check.rs
new file mode 100644
index 00000000..49359a11
--- /dev/null
+++ b/core/src/api/fact_check.rs
@@ -0,0 +1,42 @@
+// Stract is an open source web search engine.
+// Copyright (C) 2023 Stract ApS
+//
+// This program is free software: you can redistribute it and/or modify
+// it under the terms of the GNU Affero General Public License as
+// published by the Free Software Foundation, either version 3 of the
+// License, or (at your option) any later version.
+//
+// This program is distributed in the hope that it will be useful,
+// but WITHOUT ANY WARRANTY; without even the implied warranty of
+// MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
+// GNU Affero General Public License for more details.
+//
+// You should have received a copy of the GNU Affero General Public License
+// along with this program. If not, see <https://www.gnu.org/licenses/>.
+
+use std::sync::Arc;
+
+use axum::{extract, response::IntoResponse, Json};
+
+#[derive(serde::Deserialize, serde::Serialize, Debug)]
+pub struct Params {
+ claim: String,
+ evidence: String,
+}
+
+#[derive(serde::Deserialize, serde::Serialize, Debug)]
+pub struct Response {
+ score: f64,
+}
+
+pub async fn route(
+ extract::State(state): extract::State<Arc<super::State>>,
+ extract::Json(params): extract::Json<Params>,
+) -> Result<impl IntoResponse, http::StatusCode> {
+ let score = state
+ .fact_checker
+ .run(&params.claim, &params.evidence)
+ .map_err(|_| http::StatusCode::INTERNAL_SERVER_ERROR)?;
+
+ Ok(Json(Response { score }))
+}
diff --git a/core/src/api/mod.rs b/core/src/api/mod.rs
index e0814e92..56344b3a 100644
--- a/core/src/api/mod.rs
+++ b/core/src/api/mod.rs
@@ -25,6 +25,7 @@ use crate::{
cluster::Cluster,
member::{Member, Service},
},
+ fact_check_model::FactCheckModel,
improvement::{store_improvements_loop, ImprovementEvent},
leaky_queue::LeakyQueue,
qa_model::QaModel,
@@ -51,6 +52,7 @@ mod alice;
mod autosuggest;
mod chat;
mod explore;
+mod fact_check;
pub mod improvement;
mod index;
mod metrics;
@@ -71,6 +73,7 @@ pub struct State {
pub search_counter_success: crate::metrics::Counter,
pub search_counter_fail: crate::metrics::Counter,
pub summarizer: Arc<Summarizer>,
+ pub fact_checker: Arc<FactCheckModel>,
pub improvement_queue: Option<Arc<Mutex<LeakyQueue<ImprovementEvent>>>>,
pub cluster: Arc<Cluster>,
}
@@ -149,6 +152,7 @@ pub async fn router(
search_counter_fail,
remote_webgraph,
summarizer: Arc::new(Summarizer::open(&config.summarizer_path)?),
+ fact_checker: Arc::new(FactCheckModel::open(&config.fact_check_model_path)?),
improvement_queue: query_store_queue,
cluster,
});
@@ -195,7 +199,8 @@ pub async fn router(
)
.route("/beta/api/webgraph/knows_site", get(webgraph::knows_site))
.route("/beta/api/alice", get(alice::route))
- .route("/beta/api/alice/save_state", post(alice::save_state)),
+ .route("/beta/api/alice/save_state", post(alice::save_state))
+ .route("/beta/api/fact_check", post(fact_check::route)),
)
.with_state(state))
}
diff --git a/core/src/fact_check_model.rs b/core/src/fact_check_model.rs
new file mode 100644
index 00000000..e7d3f706
--- /dev/null
+++ b/core/src/fact_check_model.rs
@@ -0,0 +1,97 @@
+// Stract is an open source web search engine.
+// Copyright (C) 2023 Stract ApS
+//
+// This program is free software: you can redistribute it and/or modify
+// it under the terms of the GNU Affero General Public License as
+// published by the Free Software Foundation, either version 3 of the
+// License, or (at your option) any later version.
+//
+// This program is distributed in the hope that it will be useful,
+// but WITHOUT ANY WARRANTY; without even the implied warranty of
+// MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
+// GNU Affero General Public License for more details.
+//
+// You should have received a copy of the GNU Affero General Public License
+// along with this program. If not, see <https://www.gnu.org/licenses/>.
+
+use std::path::Path;
+
+use itertools::Itertools;
+use tch::Tensor;
+use tokenizers::PaddingParams;
+use tokenizers::TruncationParams;
+
+const TRUNCATE_INPUT: usize = 512;
+
+#[derive(thiserror::Error, Debug)]
+pub enum Error {
+ #[error("Tokenizer error")]
+ Tokenizer(#[from] tokenizers::Error),
+ #[error("Torch error")]
+ Torch(#[from] tch::TchError),
+}
+
+pub struct FactCheckModel {
+ tokenizer: tokenizers::Tokenizer,
+ model: tch::CModule,
+}
+
+impl FactCheckModel {
+ pub fn open<P: AsRef<Path>>(folder: P) -> Result<Self, Error> {
+ let truncation = TruncationParams {
+ max_length: TRUNCATE_INPUT,
+ ..Default::default()
+ };
+
+ let padding = PaddingParams {
+ ..Default::default()
+ };
+
+ let mut tokenizer =
+ tokenizers::Tokenizer::from_file(folder.as_ref().join("tokenizer.json"))?;
+
+ tokenizer.with_truncation(Some(truncation));
+ tokenizer.with_padding(Some(padding));
+
+ let model = tch::CModule::load(folder.as_ref().join("model.pt"))?;
+
+ Ok(Self { tokenizer, model })
+ }
+
+ pub fn run(&self, claim: &str, evidence: &str) -> Result<f64, Error> {
+ let encoded = self.tokenizer.encode((claim, evidence), true)?;
+
+ let ids = encoded.get_ids().iter().map(|i| *i as i64).collect_vec();
+
+ let attention_mask = encoded
+ .get_attention_mask()
+ .iter()
+ .map(|i| *i as i64)
+ .collect_vec();
+
+ let ids = Tensor::from_slice(&ids).unsqueeze(0);
+ let attention_mask = Tensor::from_slice(&attention_mask).unsqueeze(0);
+
+ let output = self.model.forward_ts(&[ids, attention_mask])?;
+
+ Ok(output.double_value(&[0, 0]))
+ }
+}
+
+#[cfg(test)]
+mod tests {
+ use super::*;
+
+ #[test]
+ fn sanity_check() {
+ let model =
+ FactCheckModel::open("../data/fact_model").expect("Failed to find fact-check model");
+
+ let score = model.run(
+ "Albert Einstein work in the field of computer science",
+ "Albert Einstein was a German-born theoretical physicist, widely acknowledged to be one of the greatest and most influential physicists of all time.",
+ ).unwrap();
+
+ assert!(score > 0.1);
+ }
+}
diff --git a/core/src/lib.rs b/core/src/lib.rs
index f00d1384..3a21e077 100644
--- a/core/src/lib.rs
+++ b/core/src/lib.rs
@@ -43,6 +43,7 @@ mod distributed;
mod entity_index;
mod enum_map;
mod executor;
+mod fact_check_model;
mod fastfield_reader;
mod human_website_annotations;
mod hyperloglog;
@@ -172,6 +173,7 @@ pub struct FrontendConfig {
pub qa_model_path: Option<String>,
pub bangs_path: String,
pub summarizer_path: String,
+ pub fact_check_model_path: String,
pub query_store_db_host: Option<String>,
pub cluster_id: String,
pub gossip_seed_nodes: Option<Vec<SocketAddr>>,
@@ -332,6 +334,9 @@ pub enum Error {
#[error("Alice")]
Alice(#[from] crate::alice::Error),
+
+ #[error("Fact check model")]
+ FactCheckModel(#[from] crate::fact_check_model::Error),
}
pub(crate) type Result<T> = std::result::Result<T, Error>;
diff --git a/frontend/src/pages/chat.astro b/frontend/src/pages/chat.astro
index e17bf14d..e4f40b46 100644
--- a/frontend/src/pages/chat.astro
+++ b/frontend/src/pages/chat.astro
@@ -145,11 +145,29 @@ import OpticSelector from "../components/OpticSelector.astro";
type Message = {
from: "user" | "alice",
message: string,
- queries: Query[]
- sourceColors: Record<number, Record<number, string>> // queryId -> sourceId -> color
+ queries: Query[],
+ sourceColors: Record<number, ScoredColor> // idx -> color
};
type MessageWithElement = Message & {element: HTMLDivElement | null};
+ type Color = {
+ r: number,
+ g: number,
+ b: number,
+ a: number
+ };
+
+ type ScoredColor = {
+ color: Color,
+ score: number
+ };
+
+ let colorGreen: Color = {r: 22, g: 163, b: 74, a: 1};
+ let colorYellow: Color = {r: 234, g: 179, b: 8, a: 1};
+ let colorRed: Color = {r: 220, g: 38, b: 38, a: 1};
+
+ let scoreGradient: Color[] = colorGradient(colorRed, colorYellow, 5).concat(colorGradient(colorYellow, colorGreen, 5));
+
let numMessageLines = 1;
let messageInputHeight = 24;
let firstMessage = true;
@@ -164,7 +182,7 @@ import OpticSelector from "../components/OpticSelector.astro";
// "from": "user",
// "message": "how does a car work",
// "queries": [],
- // "sourceColors": {},
+ // "sourceColors": [],
// "element": null
// },
// {
@@ -195,7 +213,7 @@ import OpticSelector from "../components/OpticSelector.astro";
// ]
// }
// ],
- // "sourceColors": {},
+ // "sourceColors": [],
// "element": null
// }
// ];
@@ -206,6 +224,24 @@ import OpticSelector from "../components/OpticSelector.astro";
// renderMessage(message);
// }
+ function colorGradient(start: Color, end: Color, steps: number): Color[] {
+ const gradient: Color[] = [];
+
+ for (let i = 0; i < steps; i++) {
+ const r = start.r + (end.r - start.r) * i / steps;
+ const g = start.g + (end.g - start.g) * i / steps;
+ const b = start.b + (end.b - start.b) * i / steps;
+ const a = start.a + (end.a - start.a) * i / steps;
+ gradient.push({r, g, b, a});
+ }
+
+ return gradient;
+ }
+
+ function colorToString(color: Color): string {
+ return `rgba(${color.r}, ${color.g}, ${color.b}, ${color.a})`;
+ }
+
function updateMessageInputHeight() {
const lines = messageInput.value.split("\n");
numMessageLines = lines.length;
@@ -377,7 +413,7 @@ import OpticSelector from "../components/OpticSelector.astro";
from: from,
message: message,
queries: queries,
- sourceColors: {},
+ sourceColors: [],
element: null
});
@@ -414,6 +450,7 @@ import OpticSelector from "../components/OpticSelector.astro";
let source2Id: Record<number, Record<number, number>> = {}; // (i, j) -> id
+ let sourceIdx = 0;
while ((match = regex.exec(renderedMessage)) !== null) {
const queryNum = parseInt(match[1]);
const sourceNum = parseInt(match[2]);
@@ -425,10 +462,10 @@ import OpticSelector from "../components/OpticSelector.astro";
const source = query.results[j];
// find text from '.' to match[0]
- let text = renderedMessage.substring(renderedMessage.lastIndexOf(".", match.index) + 1, match.index).trim();
+ let claim = renderedMessage.substring(renderedMessage.lastIndexOf(".", match.index) + 1, match.index).trim();
- if (text.startsWith("<p>")) {
- text = text.substring(3);
+ if (claim.startsWith("<p>")) {
+ claim = claim.substring(3);
}
@@ -441,18 +478,39 @@ import OpticSelector from "../components/OpticSelector.astro";
nextSourceId += 1;
}
+ // check if color is already assigned
+ if (!(sourceIdx in message.sourceColors)) {
+ factCheck(claim, source.text, sourceIdx).then((res) => {
+ if (res != null) {
+ let [score, sourceIdx] = res;
+ // s is a score from 0 to 1
+ let colorId = Math.floor(score * (scoreGradient.length - 1));
+ let color = scoreGradient[colorId];
+
+
+ message.sourceColors[sourceIdx] = {
+ color: color,
+ score: score
+ };
+
+ updateSourceColors(message);
+ }
+ })
+ }
+
const id = source2Id[i][j];
messageSources.push({
query: query,
source: source,
- claim: text,
+ claim: claim,
id: id
})
- const link = `<span class="inline-source"><a href="` + source.url + `">` + id + `</a></span>`;
+ const link = `<span class="inline-source inline-source-` + sourceIdx + `"><a href="` + source.url + `">` + id + `</a></span>`;
newRenderedMessage = newRenderedMessage.replace(match[0], link);
+ sourceIdx += 1;
}
// url and site of sources ordered by id
@@ -521,11 +579,53 @@ import OpticSelector from "../components/OpticSelector.astro";
message.element.appendChild(msgElem);
}
+ updateSourceColors(message);
+
// @ts-ignore
hljs.highlightAll();
removeBackgroundFromCode();
}
+ async function factCheck(claim: string, evidence: string, sourceIdx: number): Promise<[number, number] | null> {
+ let url = "/beta/api/fact_check";
+
+ const response = await fetch(url, {
+ method: "POST",
+ headers: {
+ "Content-Type": "application/json"
+ },
+ body: JSON.stringify({
+ claim: claim,
+ evidence: evidence
+ })
+ });
+
+ if (response.status == 200) {
+ const data = await response.json();
+ return [data.score, sourceIdx];
+ } else {
+ return null;
+ }
+ }
+
+ function updateSourceColors(message: MessageWithElement) {
+ if (message.element == null) {
+ return;
+ }
+
+ // loop through the sourceColors
+ for (const idx in message.sourceColors) {
+ const scoredColor = message.sourceColors[idx];
+ const source: HTMLSpanElement | null = message.element.querySelector(".inline-source-" + idx);
+
+ if (source) {
+ source.style.borderColor = colorToString(scoredColor.color);
+ source.title = Math.floor(scoredColor.score*100) + "% confidence that source supports claim";
+ }
+ }
+
+ }
+
function renderMarkdown(markdown: string) {
// @ts-ignore
const html = marked.parse(markdown);
diff --git a/justfile b/justfile
index cea9b66e..44ef975d 100644
--- a/justfile
+++ b/justfile
@@ -23,6 +23,7 @@
./scripts/export_qa_model
./scripts/export_abstractive_summary_model
./scripts/export_dual_encoder
+ ./scripts/export_fact_model
cd frontend; npm install; npm run build
cargo run --release --all-features -- configure {{ARGS}}
diff --git a/scripts/export_fact_model b/scripts/export_fact_model
new file mode 100755
index 00000000..e14dd385
--- /dev/null
+++ b/scripts/export_fact_model
@@ -0,0 +1,42 @@
+#!.venv/bin/python3
+import os
+import transformers
+import torch
+
+MODEL = "Dzeniks/roberta-fact-check"
+
+os.system("mkdir data/fact_model")
+
+model = transformers.AutoModelForSequenceClassification.from_pretrained(MODEL)
+
+model.eval()
+for p in model.parameters():
+ p.requires_grad_(False)
+
+tokenizer = transformers.AutoTokenizer.from_pretrained(MODEL)
+features = tokenizer(
+ "Albert Einstein work in the field of computer science",
+ "Albert Einstein was a German-born theoretical physicist, widely acknowledged to be one of the greatest and most influential physicists of all time.",
+ padding="max_length",
+ truncation=True,
+ return_tensors="pt",
+)
+input_ids = features["input_ids"]
+attention_mask = features["attention_mask"]
+
+tokenizer.save_pretrained("data/fact_model")
+
+
+def run_model(input_ids, attention_mask):
+ res = model(
+ input_ids=input_ids,
+ attention_mask=attention_mask,
+ return_dict=False,
+ )[0]
+ res = torch.softmax(res, dim=1)
+ return res
+
+
+traced_model = torch.jit.trace(run_model, ((input_ids, attention_mask)))
+
+traced_model.save("data/fact_model/model.pt")