Astra for Law: how legal AI search works and what Russia would need

A lawyer provides the facts of a dispute and asks for the closest precedent. General web search may return a summary, an obsolete version or a judgment reversed on appeal. Astra for Law first calls a specialist index, finds a legal authority and the relevant passage, and then applies that material to the facts.

That sequence describes the product more accurately than the model name. Legal-answer quality depends on corpus coverage, freshness, passage retrieval, the distinction between binding and persuasive sources, and citation review. The language model performs its analysis on top of that system.

According to OpenAI’s product announcement, the product combines GPT‑6 Astra with a legal search index and instructions for legal analysis and writing. The index spans more than 230 million URLs covering US cases, statutes, regulations, court rules and administrative decisions. New material is added daily.

The case-law layer uses the nonprofit Free Law Project’s CourtListener collection. The project reports coverage of more than 99.9% of published US precedential case law. Other kinds of authority also sit in the index, so 230 million URLs should not be read as 230 million cases or unique documents.

Special instructions shape the analysis: account for client facts, develop arguments on both sides, distinguish a court’s holding from secondary observations, expose weaknesses and draft possible deal terms. Instructions guide the model’s work; sources and professional legal judgment remain essential.

The query first connects the facts to legal attributes such as governing law, claim type, court, time period and disputed circumstances. Retrieval must return the right authorities and the passages that state the relevant rule. The model compares those passages with the facts, presents arguments and weaknesses, and places citations beside its claims.

In OpenAI’s published example, the system found a decision with closely matching facts and explained which differences could alter the result. The product help page explicitly tells users to review answers and cited sources before relying on them.

OpenAI has not published the storage design, document chunking method, semantic retrieval method, reranking algorithm or version-linking rules. Those components are common in modern search systems, but their implementation in Astra for Law is unknown. The public evidence establishes the product components, source categories, purpose of the index and reported evaluation.

OpenAI evaluated the complete Astra for Law configuration on 200 questions from a private validation subset of Vals AI’s Legal Research Bench. At the highest reasoning effort, the system passed the overall correctness check on 54% of questions. GPT‑6 Astra with general web search scored 38.7%. The gap is 15.3 percentage points, or about 40% relative to the baseline.

On case-law questions, the specialist system found 24% more reference cases. On an audited set of target passages, it retrieved up to 54% more relevant passages from the correct opinions. These figures support a specific technical conclusion: a specialist corpus and retrieval tool materially change the output of the same base model.

OpenAI ran the evaluation on a private US-law set, so the complete test cannot currently be reproduced independently. The overall criterion passed for 54% of answers, leaving 46% that missed at least one requirement. The result says nothing by itself about contract review, Russian law or a particular firm’s documents.

Astra for Law initially serves selected US firms through Trusted Access in ChatGPT and Codex; API access is described as coming next. OpenAI offers eligible firms Zero Data Retention on the API and excludes ChatGPT Enterprise use from human review by default.

A retention setting covers only one part of legal-system governance. A firm also needs matter-level permissions, ethical walls, audit history, deletion rules, approved sources and export controls. OpenAI says it is designing these controls with firms, while each firm remains responsible for its actual permission model and working rules.

Integrations form another layer. OpenAI launched 26 partner-built legal plugins for document, matter, discovery and time systems. iManage can save a draft to a matter file; DeepJudge can bring prior deals into a comparison. These connections supply permitted firm context and remain separate from the public legal index.

The US index does not contain the Russian authority structure a Russian lawyer needs. A domestic system can begin with the official legal publication portal, Constitutional and Supreme Court material, available general-jurisdiction decisions and the commercial-court case file. Each source requires a separate check of access method, completeness, format and reuse terms.

Every indexed document needs metadata: issuing body, instrument type, number, publication date, effective period and links to amending instruments. Court decisions need the court, instance, date, case number and review status. Semantic search without these fields may find similar wording while missing that a rule changed or a judgment was overturned.

Commercial research services and a firm’s internal collections can extend the corpus only under their licences and agreed access terms. A software project creates no right to copy a closed database. The client, its lawyers and rights holders approve the source set.

Public case law has known gaps: some case categories and protected information are withheld, published text is redacted, and general-jurisdiction search is distributed across court sites. The interface must expose those corpus limits. An empty search result does not prove that no decision exists.

An update pipeline collects permitted documents, validates their format, extracts text and splits it by legal structure: article, paragraph, clause or judicial proposition. It links versions and stores a checksum. A source change reprocesses affected records, while prior versions remain available for questions tied to an earlier date.

Retrieval combines exact identifiers and wording with semantic similarity. Filters constrain the date, issuing body, court, instance and material type. Reranking promotes passages that satisfy both the language of the question and its legal conditions. Before any generated answer, the interface can show the quotation, document version and primary-source link.

The model receives the question facts and a limited set of retrieved passages. Its answer preserves a citation for each material proposition and identifies conflicting authorities, missing facts and uncertainty. Opening a citation leads to the exact quoted location, allowing a lawyer to catch a retrieval or reasoning error before a document leaves the system.

Public law and client documents serve different purposes and belong in separate stores. A matter is visible only to its assigned team. Permission checks run before retrieval, preventing a forbidden document from entering the passage set or the model request. Encryption, access logs, retention and deletion cover source files, indexes and backups.

When collecting and subsequently processing Russian citizens’ personal data, the architecture must address Federal Law No. 152‑FZ. The law published by the Russian Ministry of Labour states in Article 18(5) that, as a general rule, databases located outside Russia may not be used during collection to record, organise, accumulate, store, clarify (update or change), or retrieve that data, subject to the exceptions listed in the law. The legal basis, operator notifications and cross-border transfer requirements must be determined for the specific process.

Article 8 of Federal Law No. 63‑FZ defines advocate-client secrecy as any information connected with an advocate’s assistance to a client. Before uploading matter material, the client must establish permitted recipients and subprocessors, retention periods, access rules and deletion procedures. A provider promise not to train on the data does not satisfy those requirements by itself.

A closed deployment can run the model and search components on client infrastructure or with a Russian provider. An external model API is suitable only where data transfer is permitted and backed by acceptable contractual controls. A data-flow diagram should identify every field leaving the system, the location of logs, key ownership and deletion procedures.

Russian lawyers create a set of real research questions and an answer key containing the relevant authorities, required passages, applicable version and grading criteria. Part of the set remains hidden from the development team. Evaluation scores document retrieval, passage retrieval, citation accuracy, application to facts and completeness of qualifications separately.

Failures are classified as well. The system may miss a document, choose weak authority, confuse versions, distort a quotation or draw a conclusion unsupported by the cited passage. That diagnosis reveals whether ingestion, filters, ranking, model instructions or the review interface needs correction.

Publication, client delivery and procedural action require a lawyer’s approval. Automated identifier and citation checks reduce mechanical mistakes, while the lawyer retains responsibility for facts, risk and legal strategy.

The studio can build the technical system: permitted-source ingestion and updates, version storage, an index with legal filters, exact-passage retrieval, a matter interface, primary-source links, roles, audit history and internal-document integrations. The model can use an approved API or run inside a closed environment.

The first production module should have a defined document set and measurable output: finding the law effective on a given date, retrieving judicial positions, reviewing a contract against an approved method, or drafting a research note with citations. Its test suite establishes minimum thresholds for retrieval, citation and access control.

The client’s legal partner defines the source hierarchy, method, permitted conclusions and approval rules. Rights holders license closed databases. pommeDeTerre is responsible for software architecture, search quality, security, interface design and reproducible evaluation. The deliverable is a reviewable system for lawyers working with Russian material.