improved chunks, added page to the llm, improved the system prompt and the markdown of the LLM response

Author: codejutsu1

app/Http/Integrations/Qdrant/Requests/QueryRequest.php

@@ -61,4 +61,30 @@ public function defaultBody(): array
 
         return $body;
     }
+
+    protected function random(): void
+    {
+        // $response = Http::withToken(
+        //     ""
+        //   )->put(
+        //     "",
+        //     [
+        //       "vectors" => [
+        //         "size" => 384,
+        //         "distance" => "Cosine"
+        //       ]
+        //     ]
+        //   );
+
+        // $response = Http::withToken(
+        //     ""
+        //   )->put(
+        //     "",
+        //     [
+        //       "field_name" => "doc_id",
+        //       "field_schema" => "keyword"
+        //     ]
+        //   );
+
+    }
 }

app/Jobs/ProcessChunkJob.php

@@ -19,9 +19,9 @@ class ProcessChunkJob implements ShouldQueue
 {
     use Batchable, Queueable;
 
-    public $tries = 6;
+    public $tries = 10;
 
-    public $backoff = 20;
+    public $backoff = 65;
 
     /**
      * Create a new job instance.
@@ -62,6 +62,7 @@ public function handle(): void
                 'page' => $this->chunk['page'] ?? null,
                 'chunk_index' => $this->chunk['chunk_index'],
                 'text' => $this->chunk['text'],
+                'pages_spanned' => $this->chunk['pages_spanned'],
             ],
         ]);
 

app/Jobs/ProcessDocumentJob.php

@@ -31,12 +31,16 @@ public function handle(): void
 
         $fileId = $file->id;
 
-        $pdfText = app(PdfService::class)->getPdfText($file->path);
+        $pdfService = app(PdfService::class);
 
-        $chunks = $this->chuckText($pdfText, 1500, 500);
+        $documentText = $pdfService->getPdfText($file->path);
+
+        $chunks = $pdfService->chunkText($documentText);
 
         $chuckCount = count($chunks);
 
+        Log::info('chunks', ['chunks' => $chuckCount]);
+
         $jobs = [];
 
         // $chunks = array_slice($chunks, 0, 2);
@@ -66,43 +70,6 @@ public function handle(): void
             ->dispatch();
     }
 
-    protected function chuckText(
-        string $text,
-        int $chunkSize = 3000,
-        int $overlap = 500
-    ): array {
-        $chunks = [];
-        $len = mb_strlen($text);
-
-        $start = 0;
-        $index = 0;
-
-        while ($start < $len) {
-            $end = min($start + $chunkSize, $len);
-            $chunk = mb_substr($text, $start, $end - $start);
-
-            $chunks[] = [
-                'text' => trim($chunk),
-                'chunk_index' => $index,
-                'char_start' => $start,
-                'char_end' => $end,
-            ];
-
-            $index++;
-
-            if ($end >= $len) {
-                break;
-            }
-
-            $start = $end - $overlap;
-            if ($start < 0) {
-                $start = 0;
-            }
-        }
-
-        return $chunks;
-    }
-
     protected function sequenceJobs(): void
     {
         // foreach ($chunks as $chunk) {

app/Jobs/ProcessUserQueryJob.php

@@ -12,11 +12,17 @@
 use Illuminate\Contracts\Queue\ShouldQueue;
 use Illuminate\Foundation\Queue\Queueable;
 use Illuminate\Support\Facades\Log;
+use Prism\Prism\ValueObjects\Messages\AssistantMessage;
+use Prism\Prism\ValueObjects\Messages\UserMessage;
 
 class ProcessUserQueryJob implements ShouldQueue
 {
     use Queueable;
 
+    public $tries = 3;
+
+    public $backoff = 10;
+
     /**
      * Create a new job instance.
      */
@@ -40,24 +46,58 @@ public function handle(): void
 
         $results = VectorDatabase::search($payload);
 
+        Log::info(['results count' => $results]);
+
         Log::info('Results', ['results' => $results]);
 
         $context = '';
 
         foreach ($results as $i => $h) {
             $p = $h['payload'];
-            $context .= "---CHUNK {$i}---\n[doc: {$p['doc_id']}, page: {$p['page']}] \n".$p['text']."\n\n";
-        }
 
-        $prompt = "\n\nContext:\n{$context}\nUser: {$this->message->message}\nAnswer:";
+            $pageLabel = '';
 
-        $llmResponse = Llm::prompt(prompt: $prompt);
+            $pages = $p['pages_spanned'] ?? [];
 
-        Log::info('LLM Response', ['llmResponse' => $llmResponse]);
+            if (is_array($pages)) {
+                if (count($pages) > 1) {
+                    $pageLabel = 'pages '.implode(', ', $pages);
+                } elseif (count($pages) === 1) {
+                    $pageLabel = 'page '.$pages[0];
+                } else {
+                    $pageLabel = 'page '.($p['page'] ?? 'N/A');
+                }
+            } else {
+                $pageLabel = 'page '.$pages;
+            }
+
+            Log::info(['pageLabel' => $pageLabel]);
+
+            $context .= "---CHUNK {$i}---\n[doc: {$p['doc_id']}, {$pageLabel}]\n".$p['text']."\n\n";
+        }
+
+        Log::info('Context', ['context' => $context]);
 
         /** @var Conversation $conversation */
         $conversation = $this->message->conversation;
 
+        $prismMessages = $conversation->messages->map(function ($message) {
+            /** @phpstan-ignore-next-line */
+            return $message->participant == MessageParticipant::USER
+                /** @phpstan-ignore-next-line */
+                ? new UserMessage($message->message)
+                /** @phpstan-ignore-next-line */
+                : new AssistantMessage($message->message);
+        })->all();
+
+        $prompt = "\n\nContext:\n{$context}\nUser: {$this->message->message}\nAnswer:";
+
+        $prismMessages[] = new UserMessage($prompt);
+
+        $llmResponse = Llm::prompt(prismMessages: $prismMessages);
+
+        Log::info('LLM Response', ['llmResponse' => $llmResponse]);
+
         $conversation->messages()->create([
             'user_id' => $this->message->user_id,
             'message' => $llmResponse,

app/Providers/AppServiceProvider.php

@@ -22,7 +22,7 @@ public function register(): void
     public function boot(): void
     {
         RateLimiter::for('pdf-processing', function () {
-            return Limit::perMinute(100);
+            return Limit::perMinute(99);
         });
     }
 }

app/Services/Llm/Driver/GeminiDriver.php

@@ -18,14 +18,12 @@ public function prompt(?array $prismMessages = [], ?User $user = null, ?string $
         try {
             return Prism::text()
                 ->using(Provider::Gemini, 'gemini-2.5-flash')
-                ->withSystemPrompt("You are a helpful AI assistant that primarily answers questions based on the legal documents provided in the Context. When a user asks a question related to the Context, use the information from it as your main source and cite each reference inline using the page id format (e.g., [Page 1]). 
-                If the user's question is general, conversational, or unrelated to the provided Context, you may respond naturally and helpfully using your own general knowledge — do NOT reply with “I couldn't find relevant information in the provided documents.”
-                Always aim to be informative, friendly, and accurate. When relevant, include citations from the Context, but when not relevant, engage normally as a helpful assistant while replying 'No relevant information found in the provided documents.'.
-                Never, never, never leave a response empty please!.")
-                // ->withMessages($prismMessages)
-                ->withPrompt($prompt)
-                ->withMaxTokens(512)
-                ->usingTemperature(0.0)
+                ->withSystemPrompt($this->getSystemPrompt())
+                ->withMessages($prismMessages)
+                // ->withPrompt($prompt)
+                ->withMaxTokens(8000)
+                ->withClientOptions(['timeout' => 60])
+                ->usingTemperature(0.2)
                 ->asText()->text;
         } catch (PrismException $e) {
             Log::error('Prompt generation failed:', [
@@ -57,4 +55,25 @@ public function embed(?string $texts = null, ?string $path = null): array
             return [];
         }
     }
+
+    protected function getSystemPrompt(): string
+    {
+        return "You are a helpful and informative AI assistant specializing in legal document analysis. Your primary goal is to answer questions accurately by prioritizing the information provided in the **Context** section below.
+
+        ### Core RAG Rules:
+        1.  **Grounded Answers:** When a user asks a question related to the Context, use only the retrieved documents as your source.
+        2.  **Citations:** For every piece of information drawn from the Context, you **must** cite the reference inline using the document and page format, such as **[page: PAGE_NUM]**.
+        3.  **Calculation Priority (New Rule):** If the user asks a question requiring arithmetic, you **must** use the tax bands and rates available in the Context to perform the calculation. State the final answer. **If specific initial income brackets are missing but subsequent brackets and rates are present, you must use reasonable assumptions derived from the common structure of the progressive tax system to complete the calculation.**
+        4.  **Non-Relevant Queries:** If the user's question is general, conversational, or unrelated to the provided Context, you must ignore the Context and respond naturally and helpfully using your own general knowledge. Do not reference the documents or apologize for not finding information.
+        5.  **No Empty Responses:** Under no circumstances should you ever return an empty response.
+
+        Always aim to be informative, friendly, and accurate. ";
+
+        // You must output the step-by-step arithmetic before stating
+        // 'You are a helpful AI assistant that primarily answers questions based on the legal documents provided in the Context. When a user asks a question related to the Context, use the information from it as your main source and cite each reference inline using the page id format (e.g., [Page 1]).
+        //         If the user's question is general, conversational, or unrelated to the provided Context, you may respond naturally and helpfully using your own general knowledge — do NOT reply with “I couldn't find relevant information in the provided documents.”
+        //         Always aim to be informative, friendly, and accurate. When relevant, include citations from the Context, but when not relevant, engage normally as a helpful assistant while replying 'No relevant information found in the provided documents.'.
+        //         Never, never, never leave a response empty please!.'
+
+    }
 }

app/Services/Pdf/PdfService.php

@@ -31,20 +31,91 @@ public function getPdf(string $filename): string
         return Storage::disk('public')->path($pdf);
     }
 
-    public function getPdfText(string $filename): string
+    public function getPdfText(string $filename): array
     {
         $parser = new Parser;
         $pdf = $parser->parseFile($this->getPdf($filename));
 
-        return $pdf->getText();
+        $pages = [];
+        foreach ($pdf->getPages() as $pageNumber => $page) {
+            $pages[] = [
+                'page' => $pageNumber + 1, // human-readable page number
+                'text' => $page->getText(),
+            ];
+        }
+
+        return $pages;
     }
 
-    // public function chuckPdfToText(
-    //     string $filename,
-    //     int $chunkSize = 3000,
-    //     int $overlap = 500,
-    // ): array
-    // {
-    //
-    // }
+    public function chunkText(
+        array $documentText,
+        int $chunkSize = 6000,
+        int $overlap = 500
+    ): array {
+        $fullText = '';
+        $pageMap = [];
+        $currentPosition = 0;
+
+        foreach ($documentText as $data) {
+            $pageMap[] = [
+                'start' => $currentPosition,
+                'page' => $data['page'],
+            ];
+
+            $fullText .= $data['text'];
+
+            $currentPosition += mb_strlen($data['text']);
+        }
+
+        $pageMap[] = ['start' => $currentPosition, 'page' => null];
+
+        $chunks = [];
+        $len = mb_strlen($fullText);
+
+        $start = 0;
+        $index = 0;
+
+        while ($start < $len) {
+            $end = min($start + $chunkSize, $len);
+            $chunk = mb_substr($fullText, $start, $end - $start);
+
+            $pagesSpanned = [];
+            $firstPage = null;
+
+            for ($i = 0; $i < count($pageMap) - 1; $i++) {
+                $pageStart = $pageMap[$i]['start'];
+                $nextPageStart = $pageMap[$i + 1]['start'];
+                $pageNumber = $pageMap[$i]['page'];
+
+                if ($start < $nextPageStart && $end > $pageStart) {
+                    if ($firstPage === null) {
+                        $firstPage = $pageNumber;
+                    }
+                    $pagesSpanned[] = $pageNumber;
+                }
+            }
+
+            $chunks[] = [
+                'text' => trim($chunk),
+                'chunk_index' => $index,
+                'char_start' => $start,
+                'char_end' => $end,
+                'page' => $firstPage,
+                'pages_spanned' => array_unique($pagesSpanned),
+            ];
+
+            $index++;
+
+            if ($end >= $len) {
+                break;
+            }
+
+            $start = $end - $overlap;
+            if ($start < 0) {
+                $start = 0;
+            }
+        }
+
+        return $chunks;
+    }
 }

app/Services/VectorDatabase/Driver/QdrantDriver.php

@@ -47,7 +47,7 @@ public function search(QdrantSearchPayload $data): array
             Log::error('Vector database search failed', [
                 'status' => $response->status(),
                 // 'body' => $response->body(),
-                // 'json' => $response->json(),
+                'json' => $response->json(),
             ]);
         }