LCA data retrieval, in a conversation.

Chinese manufacturing BOMs arrive with terms that have no direct equivalent in Ecoinvent: "背钩" (a watch case component), "IGBT模块", "502胶水". These are translation problems — previous tools left them to the practitioner. Cortex carries thousands of curated material-to-dataset mappings, built over years of LCA practitioner decisions. "IGBT模块" resolves to market for insulated gate bipolar transistor, electric vehicle application in Ecoinvent 3.12. "502胶水" maps to polyurethane adhesive — a documented proxy decision, not a guess. "绝缘栅双极晶体管" resolves differently than "IGBT模块" — same component, different naming precision, different dataset. The mapping reflects that.

The search index carries a Chinese LCA terminology dictionary spanning tens of thousands of terms, tuned specifically for LCA practice. Without it, "镍板" fragments into tokens that match everything containing "镍" — nickel battery recycling, nickel refining, nickel alloy casting. With it, "镍板" resolves to nickel plate, and the noise disappears.

Cortex searches both Chinese and English name fields simultaneously. A query for "304 不锈钢" hits the Chinese name index for HiQLCD records and the English name index for Ecoinvent and EF records in the same pass. The result table shows dataset names in both languages — "冷硬卷, 304不锈钢 (Cold-rolled annealed coil, 304 stainless steel)" — so a Chinese factory can match it to a supplier report and a German verifier can read the EU-standard name without a second lookup.

Cortex asks one question before searching when the answer would materially change the result. It does not ask otherwise.

The conditions are precise. Primary aluminum and recycled aluminum differ by roughly 10× in carbon intensity — primary in China runs 14–20 kgCO₂e/kg, recycled aluminum 0.5–1.5. A query for "铝" without a route specification produces a result set that cannot be used for either an EPD or a CBAM filing. Cortex says this explicitly before pausing: the magnitude, the two options, and a third option to search both. The interface renders clickable buttons, not a text field asking the practitioner to type an answer they may not have memorized.

For carbon steel, the BF-BOF and EAF routes differ by 3–5×. For regional electricity, the spread across grids reaches 20×. For generic terms like "adhesive" or "plastic" that map to dozens of distinct substance classes, any result without disambiguation would be noise. Cortex asks in these cases. It skips the question when the material is already specified precisely: "冷轧 304 不锈钢卷板" carries an alloy grade, a process, and a product form. No question follows.

Every clarification includes a default value and a "search both" escape. Declining advances the search; it does not abort it. Per turn, Cortex asks at most once, covering at most three dimensions in a single panel. It does not loop.

Cortex draws on fourteen authoritative LCA databases — each licensed, curated, and structured for professional LCA practice. For any given query, the relevant databases are activated based on declared use case, region, and applicable standard. The result is a single ranked candidate list, not fourteen tabs the practitioner reconciles by hand.

HiQLCD 1.4 is the China-first database: tens of thousands of China-specific LCI records spanning every Chinese provincial grid. CN-BJ and CN-SH carry different electricity carbon intensities because northwest and eastern China have different coal-to-renewables ratios. Ecoinvent's China coverage has 2.3× regional variation. HiQLCD's is 18.7×. Ecoinvent 3.12 provides international coverage and the standard naming practitioners and verifiers recognize across EU, global, and cross-border contexts. Both exist in the same query.

The rest of the fourteen: EF 3.1 for EU PEF and CBAM compliance reference data; CarbonMinds 2.0 for specialty chemicals; TianGong 1.0 for Chinese academic LCI research; WorldSteel 1.0 for industry-association steel reference data; HiQLCD-AL 2.0 for aluminum supply chain depth; Agri-footprint 7.0 for food and agriculture EPDs; and additional databases for regional and specialty coverage.

For CBAM, Cortex presents both sides in a single response. A Chinese steel exporter needs the actual production GWP from HiQLCD — the Chinese BF-BOF route at roughly 1.7–2.5 kgCO₂e/kg — and the EU reference value from EF or Ecoinvent for the declaration. Cortex retrieves both and returns them alongside each other.

Every result carries: dataset key, GWP100 (where licensed), unit, geographic region, system model, database name and version, and a link to the dataset detail page. One conversation turn. Full provenance attached.

A general AI may estimate from training data. Cortex retrieves from a licensed source. One guesses; the other knows.

Every search candidate returns a data quality indicator. The score reflects two things Cortex can compute from the record itself: version currency — whether the dataset is from the current database release or an older one — and field completeness — whether the record carries the metadata that makes it usable.

This is a structural proxy, not a five-dimension Pedigree Matrix evaluation. The Pedigree Matrix scores — Temporal, Geographic, Technology, Completeness, Reliability — are authored by human LCA practitioners when a dataset enters a source database. Cortex retrieves and displays those scores where the source database provides them. Cortex's own DQI float is a rule-engine number: 0–1, computed at query time, based on version currency (weighted 60%) and field completeness (40%).

The distinction matters. A practitioner choosing between two candidates from the same database at the same version level will see the same version-currency component; the field-completeness component will separate them if one has a defined system boundary and the other does not. The score is a tiebreaker and a filter, not a substitute for reading the dataset record.

DQI is visible in the search response and in the API. It is not a column in the exported Excel file — the export carries GWP, source, version, system model, proxy notes, and dataset links. The DQI conversation happens in the search; the export carries the evidence.

Cortex retains what matters for the next search. It does not retain the conversation itself.

After a session closes, Cortex extracts and stores: the industry and LCA purpose the user has stated, the databases they prefer, the materials they search most frequently and the process routes they've specified, and the projects they're working on. A practitioner who has said "I always use HiQLCD for Chinese projects, Ecoinvent for EU EPDs" does not say it again. A practitioner new to LCA who asked for terminology explained in plain language gets the same register in every subsequent session without re-specifying it.

The raw conversation — the actual messages — spans the active conversation window. Beyond that, the oldest turns drop. The extracted preferences do not drop. They are a user profile, not a transcript.

Say "use Ecoinvent for EU EPDs" once. Cortex applies it next session.

Data source preference changes take one confirmation. A user mid-session who says "actually, use Ecoinvent for this query" will see a one-step confirmation prompt before the session's active database switches. Confirmed, the preference is applied immediately and stored for future sessions. The user decides; Cortex records it.