refactor(ai-parser): implement chunk-based text processing for flyer parsing

- Replace fixed text window retry strategy with dynamic chunk-based approach
- Add configurable chunk size, overlap, and maximum chunks via environment variables
- Implement chunk splitting with overlap handling for context preservation
- Add chunk processing with retry logic per chunk
- Include deduplication of items across chunks
- Update logging to reflect chunk-based processing

BREAKING CHANGE: Changes the retry strategy from fixed text windows to dynamic chunk-based processing. Existing configurations may need adjustment for FLYER_AI_CHUNK_SIZE_CHARS, FLYER_AI_CHUNK_OVERLAP_CHARS, and FLYER_AI_MAX_CHUNKS.

backend/src/flyer-import/services/ai-flyer-parser.service.ts

@@ -23,6 +23,9 @@ export class AiFlyerParserService {
   private readonly logger = new Logger(AiFlyerParserService.name);
   private readonly timeoutMs: number;
   private readonly maxRetries: number;
+  private readonly chunkSizeChars: number;
+  private readonly chunkOverlapChars: number;
+  private readonly maxChunks: number;
   private mistral: any;
   private apiKey: string;
 
@@ -34,6 +37,9 @@ export class AiFlyerParserService {
 
     this.timeoutMs = this.readPositiveIntEnv('FLYER_AI_TIMEOUT_MS', 30_000);
     this.maxRetries = this.readPositiveIntEnv('FLYER_AI_RETRIES', 2);
+    this.chunkSizeChars = this.readPositiveIntEnv('FLYER_AI_CHUNK_SIZE_CHARS', 3_000);
+    this.chunkOverlapChars = this.readPositiveIntEnv('FLYER_AI_CHUNK_OVERLAP_CHARS', 300);
+    this.maxChunks = this.readPositiveIntEnv('FLYER_AI_MAX_CHUNKS', 8);
   }
 
   private async getClient(): Promise<any> {
@@ -56,56 +62,16 @@ export class AiFlyerParserService {
 
     try {
       const client = await this.getClient();
-      const textWindows = [5000, 3000, 2000];
-      const attempts = Math.max(1, Math.min(this.maxRetries + 1, textWindows.length));
-      let lastError: unknown = null;
+      const chunks = this.splitIntoChunks(text);
+      this.logger.debug(`Parsing flyer text in ${chunks.length} chunk(s)`);
 
-      for (let i = 0; i < attempts; i++) {
-        const window = textWindows[i];
-        const prompt = this.buildPrompt(text, window);
-
-        try {
-          this.logger.debug(
-            `Sending request to Mistral Tiny (attempt ${i + 1}/${attempts}, timeout=${this.timeoutMs}ms, textWindow=${window})`,
-          );
-
-          const response = await this.withTimeout<any>(
-            client.chat({
-              model: 'mistral-tiny',
-              messages: [{ role: 'user', content: prompt }],
-              temperature: 0.1,
-            }),
-            this.timeoutMs,
-            'Mistral-anrop timeout',
-          );
-
-          const content = response.choices?.[0]?.message?.content;
-          if (!content) {
-            throw new BadRequestException('Tomt svar från AI-modellen.');
-          }
-
-          this.logger.debug(`Mistral response length: ${content.length} chars`);
-
-          const jsonString = this.sanitizeJsonResponse(content);
-          const items = JSON.parse(jsonString) as Array<Record<string, unknown>>;
-
-          if (!Array.isArray(items)) {
-            throw new BadRequestException('AI returnerade inte en JSON-array.');
-          }
-
-          return items.map((item, idx) => this.normalizeAiItem(item, idx));
-        } catch (attemptErr) {
-          lastError = attemptErr;
-          if (!this.isRetryableError(attemptErr) || i === attempts - 1) {
-            throw attemptErr;
-          }
-          this.logger.warn(
-            `Mistral attempt ${i + 1} failed (${this.toErrorMessage(attemptErr)}). Retrying with shorter text window.`,
-          );
-        }
+      const allItems: AiFlyerParseResult[] = [];
+      for (let i = 0; i < chunks.length; i++) {
+        const chunkItems = await this.parseChunkWithRetry(client, chunks[i], i + 1, chunks.length);
+        allItems.push(...chunkItems);
       }
 
-      throw lastError instanceof Error ? lastError : new ServiceUnavailableException('AI-anrop misslyckades');
+      return this.dedupeItems(allItems);
     } catch (err) {
       if (err instanceof SyntaxError) {
         this.logger.error(`JSON parse error: ${String(err)}`);
@@ -255,6 +221,106 @@ Exempel på utdata:
       .trim();
   }
 
+  private splitIntoChunks(text: string): string[] {
+    const normalized = text.replace(/\r\n/g, '\n').trim();
+    if (!normalized) return [];
+
+    if (normalized.length <= this.chunkSizeChars) {
+      return [normalized];
+    }
+
+    const chunks: string[] = [];
+    let start = 0;
+    while (start < normalized.length && chunks.length < this.maxChunks) {
+      const end = Math.min(start + this.chunkSizeChars, normalized.length);
+      const chunk = normalized.slice(start, end).trim();
+      if (chunk) chunks.push(chunk);
+      if (end >= normalized.length) break;
+      start = Math.max(0, end - this.chunkOverlapChars);
+    }
+
+    return chunks;
+  }
+
+  private async parseChunkWithRetry(
+    client: any,
+    chunkText: string,
+    chunkIndex: number,
+    totalChunks: number,
+  ): Promise<AiFlyerParseResult[]> {
+    const textWindows = [3000, 2200, 1600];
+    const attempts = Math.max(1, Math.min(this.maxRetries + 1, textWindows.length));
+    let lastError: unknown = null;
+
+    for (let i = 0; i < attempts; i++) {
+      const window = textWindows[i];
+      const prompt = this.buildPrompt(chunkText, window);
+
+      try {
+        this.logger.debug(
+          `Sending request to Mistral Tiny (chunk ${chunkIndex}/${totalChunks}, attempt ${i + 1}/${attempts}, timeout=${this.timeoutMs}ms, textWindow=${window})`,
+        );
+
+        const response = await this.withTimeout<any>(
+          client.chat({
+            model: 'mistral-tiny',
+            messages: [{ role: 'user', content: prompt }],
+            temperature: 0.1,
+          }),
+          this.timeoutMs,
+          'Mistral-anrop timeout',
+        );
+
+        const content = response.choices?.[0]?.message?.content;
+        if (!content) {
+          throw new BadRequestException('Tomt svar från AI-modellen.');
+        }
+
+        this.logger.debug(`Mistral response length: ${content.length} chars`);
+
+        const jsonString = this.sanitizeJsonResponse(content);
+        const items = JSON.parse(jsonString) as Array<Record<string, unknown>>;
+
+        if (!Array.isArray(items)) {
+          throw new BadRequestException('AI returnerade inte en JSON-array.');
+        }
+
+        return items.map((item, idx) => this.normalizeAiItem(item, idx));
+      } catch (attemptErr) {
+        lastError = attemptErr;
+        if (!this.isRetryableError(attemptErr) || i === attempts - 1) {
+          throw attemptErr;
+        }
+        this.logger.warn(
+          `Mistral chunk ${chunkIndex}/${totalChunks} attempt ${i + 1} failed (${this.toErrorMessage(attemptErr)}). Retrying with shorter text window.`,
+        );
+      }
+    }
+
+    throw lastError instanceof Error
+      ? lastError
+      : new ServiceUnavailableException('AI-anrop misslyckades');
+  }
+
+  private dedupeItems(items: AiFlyerParseResult[]): AiFlyerParseResult[] {
+    const seen = new Set<string>();
+    const deduped: AiFlyerParseResult[] = [];
+
+    for (const item of items) {
+      const key = [
+        item.normalizedName,
+        item.price ?? '',
+        item.priceUnit ?? '',
+        item.offerText ?? '',
+      ].join('|');
+      if (seen.has(key)) continue;
+      seen.add(key);
+      deduped.push(item);
+    }
+
+    return deduped;
+  }
+
   private readPositiveIntEnv(key: string, fallback: number): number {
     const raw = process.env[key];
     if (!raw) return fallback;