Kien truc Production RAG Indexing - chunk, parent-child, query va summary index

Kien truc indexing cho Production RAG

Tai lieu nay tom tat cac lop indexing nen co khi dua Retrieval-Augmented Generation (RAG) vao moi truong production. Muc tieu la tang do dung cua cau tra loi, giam hallucination, cai thien toc do truy xuat va dam bao du lieu duoc kiem soat theo tenant/quyen truy cap.

1. Chunk Indexing

Tai lieu dai can duoc tach thanh cac chunk nho, co overlap vua du de giu ngu canh. Moi chunk nen luu kb_id/parent_id, tenant_id, title, source, tags, vi tri chunk va embedding rieng. Khi truy xuat, he thong co the tim tren chunk thay vi dua toan bo tai lieu vao prompt.

Ap dung:

• Tai lieu phap ly, tieu chuan, huong dan ky thuat, catalog san pham dai.
• Chunk co kich thuoc on dinh, uu tien noi dung day du y nghia.
• Can index sparse/fulltext cho title/content va semantic embedding cho chunk.

2. Parent-Child Retrieval

Parent la tai lieu goc trong Knowledge Base, child la cac sub-chunk. Truy xuat nen tim child chunk co diem cao, sau do dua ve parent context neu can them ngu canh. Cach nay giup cau tra loi vua chinh xac theo doan, vua khong mat tong the cua tai lieu goc.

Ap dung:

• Lay top child chunk de xac dinh bang chung.
• Dung metadata parent nhu source, tags, security_level, department, access_groups de kiem soat grounding va RBAC.
• Neu cau hoi can ngu canh rong, bo sung tom tat parent hoac cac chunk lan can.

3. Query Indexing / Hypothetical Questions

Moi tai lieu hoac chunk co the duoc sinh them cac cau hoi gia dinh. Khi nguoi dung hoi bang ngon ngu gan voi cau hoi gia dinh, he thong tim ra chunk lien quan nhanh hon. Day la lop index phu, khong thay the chunk embedding.

Ap dung:

• Sinh 3-8 cau hoi cho moi tai lieu quan trong.
• Luu cau hoi theo tenant_id, kb_id/chunk_id, question, embedding.
• Chi tra loi khi cau hoi duoc grounding ve source thuc trong KB.

4. Summary Indexing

Tai lieu dai nen co tom tat ngan, dung de dinh tuyen truy xuat truoc khi lay chunk chi tiet. Summary index giup giam nhieu, tang context density va ho tro reranking.

Ap dung:

• Luu summary cho parent document.
• Dung summary de chon tai lieu ung vien, sau do lay chunk chi tiet de tra loi.
• Khong dung summary lam can cu phap ly duy nhat neu cau hoi can trich dan chinh xac.

5. Hybrid Search va Two-Stage Retrieval

Production RAG nen ket hop sparse search (keyword/fulltext) voi dense search (embedding). Sau do dung reranker de sap xep lai top candidates. Luong khuyen nghi:

1. Loc theo tenant_id, quyen truy cap, module, ngon ngu va metadata.

2. Chay sparse/fulltext va semantic retrieval song song.

3. Hop nhat diem, loai trung lap theo parent/child.

4. Rerank top N bang lexical signal, semantic score, freshness, source reliability va metadata match.

5. Dua vao prompt chi cac source dat nguong grounding.

6. Metadata, RBAC va Data Governance

Moi item trong KB/index can co metadata ro rang: tenant_id, source, tags, kb_version, department, access_groups, security_level, updated_at. Retrieval khong duoc vuot qua tenant hoac quyen cua nguoi dung. Cac cau tra loi ve phap ly/tieu chuan/an toan can bat buoc co can cu tu KB hoac thong bao chua du can cu.

Nguyen tac van han:

• Metadata/RBAC la lop bao ve bat buoc, khong chi la thong tin trang tri.
• Context density quan trong hon viec nhoi nhieu tai lieu vao prompt.
• Du lieu can duoc versioning, audit log va co quy trinh cap nhat khi tai lieu het han.

Checklist san sang Production RAG

• Co bang chunk, hypothetical question, associative edge va audit theo tenant.
• KB va chunk co FULLTEXT index.
• Chunk/KB/question quan trong co embedding.
• Retrieval co hybrid search, reranking va cache namespace theo tenant.
• Cau tra loi nhay cam co grounding gate va thong bao khi KB chua du can cu.
• Admin co nut update/seed de chay lai migration, seed, chunk va xoa cache khi can.