Extract company mentions
Identify names, identifiers, countries and document context during ingestion.
COMPANY DATA FOR RAG
Company data for RAG and entity-aware retrieval.
Improve company-data retrieval in RAG applications. Resolve legal entities, attach VAT and EIN metadata, and keep company evidence linked to the right documents.
START WITHDocuments, company mentions and metadataREVIEWEntity-linked records + retrieval filters
Design the agent around evidence and policy.Define the agent host, tool scope, countries, evidence fields, exceptions and permitted actions.
Reviewed
THE PRACTICAL ANSWER
How should company data be used in a RAG system?
Retrieval-augmented generation, or RAG, supplies retrieved evidence to a language model before it answers. For company questions, resolve the legal entity before attaching documents to an index. Use country and typed company identifiers as metadata, alongside document permissions and dates. VATFind supports the company-resolution and enrichment step; your retrieval system selects the documents and generates the answer.
- A RAG system indexes contracts, filings, supplier records or company research.
- Search results confuse a parent company, subsidiary, branch or trading brand.
- Enterprise AI needs jurisdiction and identifier filters for reliable retrieval.
THE WORKFLOW
Resolve company entities before RAG retrieval
Resolve each entity
Match mentions to resolved legal-company records and preserve unresolved or ambiguous states.
Enrich the index
Attach registration number, tax ID, jurisdiction, parent context, source and timestamps as metadata.
Retrieve with entity filters
Use your governed enterprise identifier and jurisdiction context to constrain the evidence supplied to the model.
WHAT VATFIND DOES
VATFind adds a VAT and EIN identity layer before retrieval.
During ingestion, VATFind resolves company mentions and submitted VAT numbers or EINs to legal-company records. The resulting company identifiers, jurisdiction, source and retrieval time become structured metadata, helping the RAG system retrieve documents for the intended entity rather than a similar name or group company.
Evidence meaning: a verified association means the submitted VAT number or EIN is linked to the returned company record. It is not, by itself, confirmation of current tax-registration status, tax compliance or approval of the wider transaction.
For API checks, use vat_number outside the US and tax_identifier for US EINs. Optional fields can be absent when a source does not provide them.
- 01Workspace company ID
- An opaque key for linking VATFind responses inside the same workspace and environment. Use a separately governed enterprise key for long-lived cross-system joins.
- 02Entity aliases
- Legal and submitted names retained without treating every alias as a separate company.
- 03Jurisdiction metadata
- Country, region and identifier types used as retrieval filters.
- 04Company attributes
- Selected identifiers, status and address fields for enrichment and disambiguation.
- 05Provenance metadata
- Source and retrieval time attached to the entity-resolution result.
CHOOSE THE RIGHT APPROACH
Text, metadata and relationships have different roles
Searchable text
Index the document passages needed to answer a question, preserving their original source and context.
Entity metadata
Keep country, registration identifiers, source and dates as structured fields for exact filtering and citation.
Group relationships
Model established parent and subsidiary relationships separately. A shared brand should not merge documents belonging to different entities.
CONTROL OWNERSHIP
Give each team one clear responsibility.
A reliable workflow names who owns the input, who interprets the company evidence and who makes the wider decision. That separation prevents a data result from becoming an unexplained approval.
AI and search architecture
Define canonical entity IDs, indexing strategy and retrieval filters.
Enterprise data management
Resolve source records and govern legal-company attributes.
Knowledge owners
Validate document-to-entity links and maintain domain-specific exceptions.
DECISION DESIGN
Define the evidence and action at each stage.
| Stage | Evidence to retain | Controlled action |
|---|---|---|
| Entity extraction | Mention text, document context, jurisdiction and available identifiers. | Generate candidates without collapsing similarly named companies. |
| Resolution and indexing | Canonical company ID, legal name, aliases, local identifiers and confidence. | Attach documents and chunks to the resolved entity with provenance. |
| Retrieval | Entity filter, source document, date and relevant company evidence. | Retrieve for the intended entity and show uncertainty when the link is weak. |
EXAMPLE WORKFLOW
A RAG assistant searches contracts for a group brand
A contract repository contains agreements with several subsidiaries of the same group. Resolve the entity named in each agreement at ingestion, retaining unresolved mentions for review. At query time, identify whether the user means one subsidiary or the whole group. Apply that scope before sending retrieved passages to the model. Keep the document citation as well as the company-record source: the company record establishes identity, while the contract supports the claim about the agreement.
Connect the company-data toolsIMPLEMENTATION CHECKLIST
Company data RAG checklist
- Keep entity resolution separate from embedding similarity.
- Index legal names, aliases and identifiers without treating brands as companies.
- Use jurisdiction and time filters when company facts are market- or date-specific.
- Preserve document and company-source provenance through answer generation.
- Build a labelled evaluation set containing subsidiaries, translations and duplicate names.
WHAT TO MEASURE
Prove the control is useful.
Track operating quality rather than claiming that a company match guarantees the downstream outcome.
- Retrieved chunks linked to the intended legal entity
- Wrong-group-company retrieval rate
- Answers with document and entity provenance
READ THE RESULT CORRECTLY
Vector similarity is not company identity.
Embeddings can find semantically similar text but do not reliably distinguish legal entities. Entity resolution should create the identity layer, while the RAG system retains document-level permissions, citations, freshness and relevance controls.
Review data and limitationsFREQUENTLY ASKED QUESTIONS
Company data and RAG questions
Why add company IDs to a vector index?
They let the retrieval layer filter and join by legal entity instead of relying only on name similarity or document wording.
Should company data be embedded as text?
Useful descriptive fields can be embedded, but governed identifiers, country, source and dates should also remain structured metadata for exact filtering and audit.
How are subsidiaries handled?
Resolve each legal entity separately, then add group relationships as linked metadata so retrieval can include or exclude the wider group intentionally.
Can bulk feeds support index creation?
A contracted bulk snapshot or change feed can supply agreed company records for large-scale ingestion, subject to licensed fields, markets, identifiers and delivery design.
Should company identifiers be stored only inside embeddings?
No. Keep the relevant identifier, type and country in structured metadata as well. Your retrieval application can then apply explicit filters instead of relying on semantic similarity to distinguish legal entities.
Is VATFind a vector database or a RAG platform?
VATFind supplies company lookup and evidence through its documented delivery methods. Your application owns the document index, embeddings, permissions, retrieval logic and answer generation.
Sources and further reading
Microsoft: how retrieval-augmented generation uses evidence RAG retrieves relevant material and supplies it to the model as grounding data.