WorldTaxAI – Chatbot fiscal (Espagne) avec RAG & recherche hybride

Project summary

WorldTaxAI is a multilingual (EN↔ES) tax chatbot dedicated to Spanish tax law and administration. It uses a robust RAG that combines semantic search (pgvector) with lexical search (tsvector) to deliver fast, reliable, cited answers.

1. Corpus: ~60,000 documents (BOE, DGT, AEAT, PwC), unified from heterogeneous PDFs.
2. Quality: OCR, cleaning, Unicode normalization, 400–600 token chunking.
3. Indexing: PostgreSQL + pgvector (HNSW) + FTS tsvector (GIN).
4. App: FastAPI backend, React/Next front-end (streaming), chat history & deletion.
5. Multilingual: questions in English, cited results from an ES-heavy corpus.

Objectives

1. Fast access to regulations, rulings, notices, commentary.
2. Reliable, cited answers (passages + page/section deep-links).
3. Multilingual EN↔ES (terminology robust).
4. Cost control (low-cost embeddings, unified Postgres infra).
5. Chat features: history, resume, delete, export.

Why hybrid search?

1. Semantic (embeddings): captures conceptual proximity; robust to multilingual queries (EN↔ES) and paraphrases.
2. Lexical (tsvector): anchors exact references (e.g., modelo 190, art. 10, acronyms).
3. Score fusion (+ domain boosts, MMR): maximizes recall & precision without hard filters, aligned to legal intent.

Architecture & pipeline

1) Ingestion & OCR

1. Text extraction from born-digital PDFs & scans (OCR).
2. Language detection, Unicode normalization, removal of headers/footers, page numbers.

2) Structuring & chunking

1. 400–600 token segments with section path and page numbers.
2. Quality checks, deduplication, timestamps.

3) Document enrichment

1. Metadata: jurisdiction, document_type, doc_title, source_name, year.
2. Topics: primary_topic (closed list), secondary_topics (3–5 controlled tags).
3. Short abstract for previews + optional doc-level embedding.

4) Embeddings & index

1. Model: text-embedding-3-small (OpenAI).
2. Storage: PostgreSQL
   1. pgvector vector(1536)
      1. HNSW index.
3. Parallel lexical search via tsvector
   1. GIN index.

5) Hybrid retrieval & ranking

1. Vector EN↔ES
   1. exact-match lexical.
2. Normalized score fusion, domain boosts by topics/type, MMR for diversity, light re-rank.
3. Select 5–8 passages with citations (title, section, page).

6) API & Front-end

1. FastAPI: auth, sessions, pagination, purge.
2. React/Next (streaming): i18n, copy cited snippets, export conversations.
3. Speech-to-text endpoint (optional).

7) Ops & security

1. CI/CD, metrics (latency, tokens, Recall@k), budgets.
2. Privacy: API data not used to train ChatGPT; optional zero retention.
3. RLS/read-only permissions, minimal logs, per-segment traceability.

User experience

1. Per-user history: resume, rename, delete.
2. Clickable citations to the source (with context).
3. Fact-based answers: the model does not invent, it grounds in retrieved snippets.
4. Multilingual: ask in EN, results in ES, optional EN summary.
5. Streaming answers; voice input (optional).

Results & impact

1. Seconds to access previously scattered content.
2. High recall on technical queries (retención dividendos — modelo 190, establecimiento permanente, etc.).
3. Significant time savings for analysts & attorneys.
4. Ready foundation for monitoring & analytics (trends, clustering, dedup detection).

Tech stack

1. Backend: Python, FastAPI, SQLAlchemy.
2. Database: PostgreSQL, pgvector, tsvector, Docker.
3. LLM & embeddings: OpenAI (chat gpt-4o-mini, embeddings text-embedding-3-small).
4. Front-end: React/Next, streaming, i18n.
5. Ops: CI/CD, monitoring, budgets.

Timeline & budget

1. Phase 1 — scraping & preparation: delivered, ~2–3 weeks.
2. Phase 2 — RAG + API + front + deployment: ~4 weeks.
3. Phase 2 quote: €6,600 excl. VAT (21 days @ €300/day).
4. Recurring API costs: low (one-shot indexing, lightweight queries).

Client testimonial

FAQ

Does the chatbot “learn” from our data?

No. API data is not used to train ChatGPT. Embeddings and texts remain private.

Can I ask in English if the corpus is in Spanish?

Yes. Multilingual embeddings + bilingual FTS (language routing).

How do you ensure quality?

Cited answers; continuous evaluation (Recall@k, Precision@k, Faithfulness).

How does it scale?

pgvector handles hundreds of thousands of vectors. HNSW/IVF indexes for speed. Stateless API.

Security & compliance?

Private PostgreSQL, read-only permissions (RLS), history deletion, auditable CI/CD, minimal logs.

Resources & next steps

Sample technical queries

1. “Retención sobre dividendos no residentes — modelo 190 (EN→ES)”
2. “DGT criteria on digital permanent establishment (EN→ES)”
3. “Double tax relief exemption art. 10 — BOE 202x (ES)”

Product extension ideas

1. Stronger re-ranking (cross-encoder), query rewriting (expansion).
2. Quality dashboard (coverage, freshness, terminology drift).
3. Monitoring: incremental ingestion (BOE/DGT/AEAT webhooks), topic alerts.