llm_chain/chains/
map_reduce.rs

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
//! The `map_reduce` module contains the `Chain` struct, which represents a map-reduce chain.
//!
//! A map-reduce chain is a combination of two steps - a `map` step and a `reduce` step.
//! The `map` step processes each input document and the `reduce` step combines the results
//! of the `map` step into a single output.
//!
//! The `Chain` struct is generic over the type of the `Step` and provides a convenient way
//! to execute map-reduce operations using a provided `Executor`.

use crate::traits::ExecutorError;
use crate::{
    frame::Frame, output::Output, prompt::Data, serialization::StorableEntity, step::Step, tokens,
    tokens::PromptTokensError, traits::Executor, Parameters,
};
use futures::future::join_all;
use futures::future::FutureExt;
use serde::Deserialize;
use serde::Serialize;

use thiserror::Error;

/// The `MapReduceChainError` enum represents errors that can occur when executing a map-reduce chain.
#[derive(Error, Debug)]
pub enum MapReduceChainError {
    /// An error relating to the operation of the Executor.
    #[error("FormatAndExecuteError: {0}")]
    FormatAndExecuteError(#[from] crate::frame::FormatAndExecuteError),
    /// An error relating to tokenizing the inputs.
    #[error("TokenizeError: {0}")]
    TokenizeError(#[from] crate::tokens::PromptTokensError),
    #[error("The vector of input documents was empty")]
    InputEmpty,
    #[error("Error templating: {0}")]
    StringTemplate(#[from] crate::prompt::StringTemplateError),
}

/// The `Chain` struct represents a map-reduce chain, consisting of a `map` step and a `reduce` step.
///
/// The struct is generic over the type of the `Step` and provides methods for constructing and
/// executing the chain using a given `Executor`.
#[derive(Serialize, Deserialize)]
pub struct Chain {
    map: Step,
    reduce: Step,
}

impl Chain {
    /// Constructs a new `Chain` with the given `map` and `reduce` steps.
    ///
    /// The `new` function takes two instances of `Step` and returns a new `Chain` instance.
    pub fn new(map: Step, reduce: Step) -> Chain {
        Chain { map, reduce }
    }

    /// Executes the map-reduce chain using the provided `Executor`.
    ///
    /// The `run` function takes a vector of input documents, a base set of parameters, and a reference
    /// to an `Executor`. It processes the input documents using the `map` step and the `reduce` step,
    /// and returns the result as an `Option<E::Output>`.
    ///
    /// The function is asynchronous and must be awaited.
    pub async fn run<E: Executor>(
        &self,
        documents: Vec<Parameters>,
        base_parameters: Parameters,
        executor: &E,
    ) -> Result<Output, MapReduceChainError> {
        if documents.is_empty() {
            return Err(MapReduceChainError::InputEmpty);
        }
        let map_frame = Frame::new(executor, &self.map);
        let reduce_frame = Frame::new(executor, &self.reduce);

        let chunked_docs = self.chunk_documents(
            documents.clone(),
            base_parameters.clone(),
            executor,
            &self.map,
        )?;

        // Execute the `map` step for each document, combining the base parameters with each document's parameters.
        let chunked_docs_with_base_parameters: Vec<_> = chunked_docs
            .iter()
            .map(|doc| base_parameters.combine(doc))
            .collect();
        let mapped_documents: Vec<_> = join_all(
            chunked_docs_with_base_parameters
                .iter()
                .map(|doc| map_frame.format_and_execute(doc))
                .collect::<Vec<_>>(),
        )
        .await;
        let mapped_documents = mapped_documents
            .into_iter()
            .collect::<Result<Vec<Output>, _>>()?;
        let mapped_documents: Vec<Result<Data<String>, ExecutorError>> = join_all(
            mapped_documents
                .into_iter()
                .map(|x| x.to_immediate().map(|x| x.map(|y| y.as_content())))
                .collect::<Vec<_>>(),
        )
        .await;
        let mapped_documents: Vec<Data<String>> = mapped_documents
            .into_iter()
            .collect::<Result<Vec<Data<String>>, ExecutorError>>()
            .map_err(|e| {
                MapReduceChainError::FormatAndExecuteError(
                    crate::frame::FormatAndExecuteError::Execute(e),
                )
            })?;

        let mut documents = self
            .combine_documents_up_to(executor, mapped_documents, &base_parameters)
            .await?;

        if documents.is_empty() {
            return Err(MapReduceChainError::InputEmpty);
        }

        loop {
            let tasks: Vec<_> = documents
                .iter()
                .map(|doc| base_parameters.with_text(doc))
                .collect();
            let futures = tasks.iter().map(|p| reduce_frame.format_and_execute(p));
            let new_docs = join_all(futures).await;
            let new_docs = new_docs.into_iter().collect::<Result<Vec<_>, _>>()?;
            let new_docs = join_all(
                new_docs
                    .into_iter()
                    .map(|x| x.to_immediate().map(|x| x.map(|y| y.as_content()))),
            )
            .await;
            let new_docs = new_docs
                .into_iter()
                .collect::<Result<Vec<Data<String>>, ExecutorError>>()
                .map_err(|e| {
                    MapReduceChainError::FormatAndExecuteError(
                        crate::frame::FormatAndExecuteError::Execute(e),
                    )
                })?;
            let n_new_docs = new_docs.len();
            if n_new_docs == 1 {
                return Ok(Output::new_immediate(new_docs[0].clone()));
            }
            documents = self
                .combine_documents_up_to(executor, new_docs, &base_parameters)
                .await?;
        }
    }

    async fn combine_documents_up_to<E: Executor>(
        &self,
        executor: &E,
        mut v: Vec<Data<String>>,
        parameters: &Parameters,
    ) -> Result<Vec<String>, MapReduceChainError> {
        let mut new_outputs = Vec::new();
        while let Some(current) = v.pop() {
            let mut current_doc = current.extract_last_body().cloned().unwrap_or_default();
            while let Some(next) = v.last() {
                let Some(next_doc_content) = next.extract_last_body() else {
                    continue;
                };
                let mut new_doc = current_doc.clone();
                new_doc.push('\n');
                new_doc.push_str(next_doc_content);

                let params = parameters.with_text(new_doc.clone());
                let prompt = self.reduce.format(&params)?;
                let count = executor.tokens_used(self.reduce.options(), &prompt)?;
                if count.has_tokens_remaining() {
                    current_doc = new_doc;
                    v.pop();
                } else {
                    break;
                }
            }
            new_outputs.push(current_doc);
        }
        Ok(new_outputs)
    }

    fn chunk_documents<'a, E>(
        &self,
        v: Vec<Parameters>,
        base_parameters: Parameters,
        executor: &E,
        step: &Step,
    ) -> Result<Vec<Parameters>, PromptTokensError>
    where
        E: Executor + 'a,
    {
        let data: Result<Vec<_>, _> = v
            .iter()
            .map(|x| {
                <E as tokens::ExecutorTokenCountExt>::split_to_fit(
                    executor,
                    step,
                    x,
                    &base_parameters,
                    None,
                )
            })
            .collect();
        let data = data?.iter().flatten().cloned().collect();
        Ok(data)
    }
}

/// Implements the `StorableEntity` trait for the `Chain` struct.
///
/// This implementation provides a method for extracting metadata from a `Chain` instance, in order to identify it
impl StorableEntity for Chain {
    /// Returns metadata about the Chain instance.
    ///
    /// The metadata is returned as a vector of tuples, where each tuple contains a key-value pair
    /// representing a metadata field and its value.
    ///
    /// This method also extracts metadata from the Step instances associated with the Chain.
    fn get_metadata() -> Vec<(String, String)> {
        let base = vec![(
            "chain-type".to_string(),
            "llm-chain::chains::map_reduce::Chain".to_string(),
        )];
        base
    }
}