Multilingual RAG Knowledge Agent for a Japanese Manufacturing Company

The Query Planner (gpt-4o-mini, around 300 milliseconds) forms the first stage. It knows the domain, brand codes, product types, partner names, typical sheet labels, and breaks every question down into a JSON plan: a semantic query in the original language for embedding recall, multilingual keyword chains (Japanese + English + Thai) for full-text search, automatically detected product codes and brand aliases as hard filters, predicted Excel sheet names and a complexity tag that assigns the question to one of the four answer tiers. An employee resolver translates names from the question, including romanized forms such as "mori-san", into the kanji spelling of the document corpus, and a Redis cache answers repeated phrasings in under 30 milliseconds.

Parallel Discovery forms the heart of the system: pgvector with an HNSW index delivers semantic similarity (1,536-D OpenAI embeddings, cosine), PGroonga with a TokenBigram tokenizer finds exact matches in Japanese text, LightRAG supplies entity relationships from the knowledge graph, and depending on the detected intent, tabular, email, entity-fanout and Kanban search are added. All channels run in parallel, the total latency is that of the slowest channel, not the sum. Reciprocal Rank Fusion (k=60, sparse 0.6 / dense 0.4) consolidates the lists with source-diversity enforcement, which prevents a single document from dominating the top results. HyDE, a hypothetical answer as an additional search vector, and a JP↔EN synonym dictionary with over 40 term groups close the gap between the languages of the corpus.

The hybrid search with Japanese kanji was the biggest technical challenge. Standard full-text search fails on Japanese: no spaces between words, multiple readings per character, compound terms. The solution: PGroonga splits Japanese text into overlapping 2-character fragments, bigrams. A search for "品質管理" thus also finds "品質管理報告書", without any morphological analysis ever taking place. Combined with a Supabase function that automatically resolves Japanese particles into OR conjunctions, the system achieves a hit rate that surprises even experienced Japanese users.

The document ingestion was a challenge in its own right, and today draws on four sources: SharePoint via Microsoft Graph delta sync (including mirrored document permissions, every user only sees what they would be allowed to see in SharePoint), email mailboxes with a two-stage relevance filter, meeting transcripts via webhook and browser upload. Mistral OCR parses PDFs up to 35 MB (including scanned ones), ExcelJS handles XLSX, Unstructured.io DOCX and PPTX. Japanese texts require a language-dependent chunking strategy: target 800 characters for Japanese, 1,200 for Latin script, overlap 150, with sentence-boundary detection at the Japanese full stop 「。」 and paragraph breaks in English. A SHA-256-based deduplication across the file and chunk level prevents redundant processing on repeated uploads.

Excel is the lingua franca of manufacturing, and the hardest nut to crack for any RAG. A question like “What is the current unit price for item X?” targets a single cell in a sheet with dozens of tabs, hundreds of rows and a header logic that nobody ever standardized.

The Tabular Deep Dive logic solves this in two steps. During ingestion, the system stores two representations per row: a tab-separated plain text for the PGroonga search and an HTML table for the later LLM synthesis. On a query, the system retrieves not only the matching row but also the header rows, seed rows and sibling rows of the same table block, so the model sees the full context, not just a single cell.

Before synthesis, the HTML tables are converted into compact Markdown (around 30 % token savings), and a tier-based context budget between 15,000 and 130,000 characters ensures the model always sees complete sources, three full documents rather than ten truncated ones. Each source is passed with its document name, update date and web URL, so the answer ultimately contains not just the figure, but also the reference: “[Cost Calculation Q1.xlsx] as of 2025-03-15”.

Where vector and full-text search end at individual documents, the LightRAG knowledge graph begins. During import, it automatically extracts entities, sites, suppliers, product lines, reports, and relations between them: who supplies what, who reviews what, who documents what. For cross-document questions, it provides a short path through the graph so that the synthesis understands the connection rather than guessing it. The visualization below shows a simplified view of a manufacturing domain, the lines are relationships that the graph learned from the corpus itself.

The orchestration of five AI providers, Anthropic, OpenAI, Cohere, Mistral, Unstructured.io, was a deliberate architectural decision. Each has a clearly defined role: precise multilingual embeddings, OCR quality on scanned PDFs, semantic re-ranking, cost-efficient synthesis and Office document extraction. Instead of answering every request with the most expensive model, the planner classifies each question into one of four tiers, Flash (≤4 s), Fast (≤8 s) and Standard (≤18 s) with Claude Haiku 4.5, Deep (≤35 s) with Claude Sonnet 4.6 for multi-document analyses. An Anti-Loop Session State tracks chunks already seen and prevents the agent from repeating the same results on follow-up questions; guardrails block sensitive topics such as salaries or personnel decisions. Operationally, circuit breakers, exponential retry with rate-limit awareness, a cost tracker with hourly and daily budgets per provider and an alert manager safeguard stability in production.