How AI Builds an EB1A Client Knowledge Base

After AI document classification sorts and maps a client's 30–200 uploaded documents, the next step is knowledge base construction — building the compact, structured representation of everything relevant to the petition that will power petition generation. The knowledge base is the bridge between raw documents and petition-quality prose. Under 8 CFR 204.5(h)(3), a petitioner must demonstrate compliance with each criterion through documentary evidence — the KB structures that evidence for AI-assisted generation while keeping the attorney in control of what gets argued and how. Without this structure, AI-assisted drafting either requires passing enormous volumes of raw text (context-inefficient, noise-heavy) or defaults to generic templates that don't reflect the client's actual evidence record.

What a Knowledge Base Contains

The EB1A client knowledge base is a structured document containing all the information the petition generator needs to draft every section of the petition letter. It contains five categories of content:

Client profile. Name, field of endeavor, career summary, current employer and role, petition strategy summary, and the primary criteria being argued. This section provides the generation context for the opening section of the petition letter and ensures that every generated section is grounded in the client's specific field and career narrative.

Per-criterion evidence inventories. For each of the 10 EB1A criteria — whether being argued or not — a summary of what evidence exists, how strong it is, and which documents support it. Criteria with strong evidence are flagged for detailed argumentation; criteria with weak or missing evidence are flagged for attorney attention before generation begins.

Key facts with exhibit references. Each fact needed for petition drafting is stored as a discrete entry: "Won the IEEE Best Paper Award at ICLR 2023 (Exhibit 1, Exhibit 2)." Each entry includes the exhibit number(s) that document the fact, so the petition generator can cite the correct exhibit in the generated text.

Evidence gaps and strategic notes. Attorney annotations about which evidence is strongest, how to frame specific criterion arguments, and what context is not documented in the record but should be added to the petition brief. This is the layer of attorney judgment that distinguishes an AI-assisted petition from a fully automated one.

Biographical and career timeline. The chronological record of the client's professional history — education, positions, publications, awards, contributions, media coverage — in date order. This feeds the career summary section and supports the sustained acclaim argument at Step 2.

Why Compact Representation Matters

The naive approach to AI-assisted petition drafting passes all client documents to the AI and asks it to write the petition. This fails in practice for three reasons:

Context window constraints. Even with large-context models, 200 client documents containing 500,000–2,000,000 words exceeds the volume of information the model can reason about precisely. Models working with extremely large contexts produce less focused, less accurate output than models working with compact, curated inputs.

Signal-to-noise ratio. The Criterion 5 section of the petition needs the expert letters, the publication abstracts, and the contribution evidence — not the 40-page salary history from 2018 or the childhood achievement certificate. When all documents are passed simultaneously, relevant facts are buried in noise, and the generator is as likely to surface irrelevant information as relevant information.

Reproducibility and consistency. A structured KB produces consistent output across multiple generation runs. Passing raw documents produces variable results depending on which parts of the context the model attends to in each run. Petition generation should be deterministic enough that the attorney can regenerate a section after editing without getting a completely different result.

How the KB Is Built

The KB construction process flows from document classification:

Step 1: Classification. Each document is identified by type and mapped to the criteria it supports. See how AI classifies EB1A supporting documents for the full classification pipeline. The classification results determine how each document is processed in the next step.

Step 2: Fact extraction. The AI reads each classified document and extracts the key facts relevant to the petition. For an award certificate: award name, awarding organization, year, stated selection criteria, and any mention of prior recipients. For a publication: title, journal, authors, year, citation count if mentioned. For an expert letter: the signatory's credentials, the specific contribution addressed, the significance claims made.

Step 3: Criterion organization. Extracted facts are compiled into criterion-level summaries. "Evidence for Criterion 1 includes: [Award A from Organization B in Year C based on X criteria]; [Award D from Organization E in Year F, including prior recipients Y and Z]..."

Step 4: KB assembly. The client profile, criterion summaries, and fact entries with exhibit references are assembled into the complete KB structure.

Step 5: Attorney review. The attorney reviews the KB for factual accuracy, corrects errors, adds contextual notes, and flags evidence that should be treated strategically. This review typically takes 30–60 minutes for a 50–100 document intake.

How the KB Feeds Petition Generation

When generating each petition section, the generation model receives three inputs:

The complete KB as context. The entire structured KB — client profile, criterion summaries, key facts with exhibit references — is passed to the generation model. Because the KB is compact (10,000–50,000 words rather than 2,000,000), it fits comfortably in context and is available for the model to reason about precisely.

Section-specific instructions. Instructions for the specific section being generated: which criterion to argue, which regulatory standard to address, which evidence to feature. These instructions reference specific KB entries that should anchor the generated text.

Style and structure guidance. The target structure for the section (regulatory text, evidence discussion, significance argument, exhibit citations), the tone (formal legal prose), and citation format conventions (parenthetical exhibit references).

The generated text cites KB facts using exhibit references — each claim in the generated petition can be traced back to a KB entry and from there to the source document. This traceability is the mechanism that prevents hallucinations: the generator uses facts from the KB, not from its general training knowledge about immigration petitions.

Editing and Enriching the KB

The AI-generated KB is the starting point, not the final product. Attorneys add value at several levels:

Correct factual errors. The AI may extract a publication year incorrectly, or misread the awarding organization's name. Attorneys correct these before generation to prevent incorrect facts from propagating into the petition letter.

Add context documents don't contain. An award certificate says nothing about the award's prestige relative to other awards in the field. The attorney who knows that this award is presented to only three researchers per year, selected from nominations across 50 countries, adds this context as a KB annotation. This context feeds directly into the petition narrative.

Mark strategic emphasis. Some documents support multiple criteria but should be featured prominently in one. The attorney marks which criterion to lead with for each multi-use document, and the generation system emphasizes accordingly.

Flag gaps for resolution before filing. If the KB reveals that a criterion is being argued without strong evidence — for example, the attorney planned to use Criterion 8 but the client's salary documentation only reaches the 70th percentile, not the 90th — this gap is visible in the KB before a word of the petition letter is drafted. Better to address the gap now than to generate a weak criterion argument and then have to revise.

The KB in Multi-Case Practice

For attorneys handling multiple concurrent EB1A cases, the KB architecture creates operational leverage. Each case's KB is a complete, self-contained representation of that client's record — an attorney reviewing a case after a month away can get back up to speed in 15 minutes by reading the KB, rather than re-reading 150 documents. Associates working on the same case can start from the KB rather than re-doing document review.

The KB also creates a documentation record for the case file. If the case is transferred to another attorney, or if the client needs a motion to reopen filed after a denial, the KB provides an organized case history that would otherwise require reconstruction from the raw document file.

For how documents are classified before KB construction, see how AI classifies EB1A supporting documents. For how the KB feeds into the petition generation pipeline, see how RAG powers EB1A petition drafting. For the complete end-to-end preparation workflow and how these components fit together, see the EB1A petition guide and EB1A drafting efficiency: from 200 hours to 40. The USCIS Policy Manual, Volume 6, Part F, Chapter 2 defines the adjudication standards that drive KB structure — each criterion section in the KB directly maps to a criterion the adjudicator will evaluate.

Immigration Copilot builds, stores, and serves the client knowledge base as part of the full petition preparation pipeline. Get started →