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
|
import { Networking } from '../../../../Network';
import { Observation, RAGChunk } from '../types/types';
import { ParametersType } from './ToolTypes';
import { Vectorstore } from '../vectorstore/Vectorstore';
import { BaseTool } from './BaseTool';
const ragToolParams = [
{
name: 'hypothetical_document_chunk',
type: 'string',
description: "A detailed prompt representing an ideal chunk to embed and compare against document vectors to retrieve the most relevant content for answering the user's query.",
required: true,
},
] as const;
type RAGToolParamsType = typeof ragToolParams;
export class RAGTool extends BaseTool<RAGToolParamsType> {
constructor(private vectorstore: Vectorstore) {
super(
'rag',
'Perform a RAG search on user documents',
ragToolParams,
`
When using the RAG tool, the structure must adhere to the format described in the ReAct prompt. Below are additional guidelines specifically for RAG-based responses:
1. **Grounded Text Guidelines**:
- Each <grounded_text> tag must correspond to exactly one citation, ensuring a one-to-one relationship.
- Always cite a **subset** of the chunk, never the full text. The citation should be as short as possible while providing the relevant information (typically one to two sentences).
- Do not paraphrase the chunk text in the citation; use the original subset directly from the chunk.
- If multiple citations are needed for different sections of the response, create new <grounded_text> tags for each.
2. **Citation Guidelines**:
- The citation must include only the relevant excerpt from the chunk being referenced.
- Use unique citation indices and reference the chunk_id for the source of the information.
- For text chunks, the citation content must reflect the **exact subset** of the original chunk that is relevant to the grounded_text tag.
**Example**:
<answer>
<grounded_text citation_index="1">
Artificial Intelligence is revolutionizing various sectors, with healthcare seeing transformations in diagnosis and treatment planning.
</grounded_text>
<grounded_text citation_index="2">
Based on recent data, AI has drastically improved mammogram analysis, achieving 99% accuracy at a rate 30 times faster than human radiologists.
</grounded_text>
<citations>
<citation index="1" chunk_id="abc123" type="text">Artificial Intelligence is revolutionizing various industries, especially in healthcare.</citation>
<citation index="2" chunk_id="abc124" type="table"></citation>
</citations>
<follow_up_questions>
<question>How can AI enhance patient outcomes in fields outside radiology?</question>
<question>What are the challenges in implementing AI systems across different hospitals?</question>
<question>How might AI-driven advancements impact healthcare costs?</question>
</follow_up_questions>
</answer>
`,
`Performs a RAG (Retrieval-Augmented Generation) search on user documents and returns a set of document chunks (text or images) to provide a grounded response based on user documents.`
);
}
async execute(args: ParametersType<RAGToolParamsType>): Promise<Observation[]> {
const relevantChunks = await this.vectorstore.retrieve(args.hypothetical_document_chunk);
const formattedChunks = await this.getFormattedChunks(relevantChunks);
return formattedChunks;
}
async getFormattedChunks(relevantChunks: RAGChunk[]): Promise<Observation[]> {
try {
const { formattedChunks } = await Networking.PostToServer('/formatChunks', { relevantChunks });
if (!formattedChunks) {
throw new Error('Failed to format chunks');
}
return formattedChunks;
} catch (error) {
console.error('Error formatting chunks:', error);
throw error;
}
}
}
|
