[Security GenAI][BUG] KB index entry created via pdf upload does not …

…give the right response (#198020)

## Summary

These changes fix the issue with the wrong response of the AI Assistant
using knowledge base tool and index entry generated from a PDF file.

The issue happens because we are using the first chunk of uploaded PDF
document as a context that we pass to LLM instead of using inner hits
chunks which are actual parts of the document relevant to the questions.

Here is [the blog
post](https://www.elastic.co/search-labs/blog/semantic-text-with-amazon-bedrock)
that talks about the strategy of using inner hits to get the most
relevant documents. (see `Strategy 1: API Calls` section)

### Upload + index PDF

1. Navigate to Integrations page
2. Select "Upload a file"
3. Select and upload a PDF file
4. Press Import button
5. Switch to Advanced tab
6. Fill in "Index name"
7. Add additional field > Add semantic text field > Fill in form
  * Field: `attachment.content`
  * Copy to field: `content`
  * Inference endpoint: `elser_model_2`
8. Press Add button
9. Press Import button

### Add KB index entry (with uploaded PDF data)

1. Navigate to AI Assistant's Knowledge Base page
2. New > Index
3. Fill in "New index entry" form (below are main fields)
  * Name: `[add entry name]`
  * Index: `[select index name created during uploading a PDF file]`
  * Field: `content`
4. Press Save button

### Testing notes

Enable knowledge base feature via

```
xpack.securitySolution.enableExperimental:
  - 'assistantKnowledgeBaseByDefault'
```

### Example PDF for testing

**PDF document**:
[Elastic Global Threat Report
2024](https://github.com/user-attachments/files/17544720/elastic-global-threat-report-2024.pdf)

**KB Index entry**:
Data Description: "Use this tool to answer questions about the Elastic
Global Threat Report (GTR) 2024"
Query Instruction: "Key terms to return data relevant to the Elastic
Global Threat Report (GTR) 2024"

**Questions**:
1. Who are the authors of the GTR 2024?
2. What is the forecast for the coming year in GTR 2024?
3. What are top 10 Process Injection by rules in Windows endpoints in
GTR 2024?
4. What is the most widely adopted cloud service provider this year
according to GTR 2024?
6. Give a brief conclusion of the GTR 2024

**Current behaviour**:

<img width="656" alt="Screenshot 2024-10-28 at 16 43 48"
src="https://github.com/user-attachments/assets/90615356-8807-4786-b58d-ca28c83aaec9">

**Fixed behaviour**:

<img width="655" alt="Screenshot 2024-10-28 at 16 44 47"
src="https://github.com/user-attachments/assets/9ebefbcc-20c2-4c79-98f3-11fa6acf3da6">

x-pack/plugins/elastic_assistant/server/ai_assistant_data_clients/knowledge_base/helpers.ts

@@ -220,6 +220,17 @@ export const getStructuredToolForIndexEntry = ({
               return { ...prev, [field]: hit._source[field] };
             }, {});
           }
+
+          // We want to send relevant inner hits (chunks) to the LLM as a context
+          const innerHitPath = `${indexEntry.name}.${indexEntry.field}`;
+          if (hit.inner_hits?.[innerHitPath]) {
+            return {
+              text: hit.inner_hits[innerHitPath].hits.hits
+                .map((innerHit) => innerHit._source.text)
+                .join('\n --- \n'),
+            };
+          }
+
           return {
             text: get(hit._source, `${indexEntry.field}.inference.chunks[0].text`),
           };