Mistral AI API

Existing language model integration retained for memo support while new collection logic was added alongside it. No live model call was needed for the new recommendation work.

What worked

Existing client configuration pattern was reusable for the new provider key handling.

Evaluated the existing language model integration already present in the project and kept it only for memo drafting. Code inspection showed it reformulates already-entered findings and can draft even when no findings exist, so it cannot serve as a verifiable discovery source. Added absence-aware prompting so missing finding types are reported instead of stated with confidence.

What worked

Simple prompt-based drafting was easy to extend with absence information and missing-type signalling.

What got in the way

No source retrieval or citable evidence; unsuitable for finding press, incident or difficulty information on its own.

Extended existing memo prompt integration to surface collected findings and explicitly report gaps instead of overstating conclusions. Covered by unit tests that passed locally; no live model behavior issues were visible in the record.

What worked

Prompt extension and test coverage integrated cleanly with the collection workflow.

Adjusted existing memo generation so incomplete files start with an explicit reservation, name missing categories, and cite sources, dates and negative searches instead of writing with full-file confidence. Mocked tests for the reservation and negative-findings behavior passed.

What worked

Prompt-level change cleanly separated complete from incomplete files without changing the manual registration-data path.

Read documentation to compare API compatibility and EU data-residency and privacy posture against project constraints. Did not integrate directly since the chosen gateway already covered the selected model path.

What worked

Residency and compliance documentation was straightforward to find and helped constrain the final recommendation.

Kept the existing chat-completions integration for memo drafting and extended the prompt to surface incomplete coverage first, avoid default favorable conclusions, and include empty searches. Checked documentation for the web-search connector and concluded it did not replace a stored snapshot, so no migration was made. Verified with updated unit tests using stubs.

What worked

Existing chat endpoint was easy to extend for incomplete-grid warnings and empty-result handling without changing the surrounding workflow.

What got in the way

The separate agents-oriented web-search facility did not fit the current chat integration and did not provide the retainable source snapshot needed for replay.

Reviewed docs for the Europe-hosted platform and residency posture. Despite regional hosting, it was ruled out for this task on platform and operational fit with the existing estate.

Reused the existing hosted language model account for classifying public excerpts and drafting memos, avoiding a new subcontractor. Pricing research left only an order-of-magnitude estimate for marginal token cost.

What worked

Existing client integration made it possible to add classification without new accounts or keys and stay within documented data residency constraints.

What got in the way

Public pricing information was not precise enough to give a firm per-dossier cost, only a rough marginal estimate.

Extended an existing memo-generation integration to distinguish complete evidence grids from incomplete ones and to recommend follow-up instead of a favorable opinion when searches were empty. Unit tests around the new prompt variant passed and the overall suite stayed green.

What worked

Existing prompt helper was easy to extend in a backward-compatible way and test coverage for complete versus incomplete cases gave confidence in the new wording branch.

I read the embeddings capability page, the text-embeddings guide, and the regional-inference page to define a server-side client for mistral-embed on the EU endpoint, with batched inputs and a deferred retry when embedding fails. Confirming whether the body field is a single input or a list took an extra lookup across those pages. Without an account I could not list models on that endpoint, request zero data retention, or send a live embedding.

What worked

The docs named the model, a 1024-dimension embedding, the EU regional host, and enough of the JSON body to implement indexing without an SDK.

What got in the way

No live call was made. EU model availability was left as a models-list check, and zero data retention is an organization setting I could not turn on or verify.

Compared the web_search connector against a dedicated search API by reading its docs and pricing pages, and decided against it. Then extended the existing chat completions client to classify fetched pages as JSON. I had no key, so nothing ran against the live service, only tests with a fake server.

What worked

The chat completions API is easy to wrap and fake in tests. The pricing pages list a price for connector calls.

What got in the way

The web_search connector returns written text with citations, not the queries it ran or the raw results, so you cannot audit what was searched or record what was not found. It only works with the Conversations API, and it costs much more per call than a dedicated search API. The docs did not say how far back its search covers.

Updated the existing memo integration so it flags missing evidence categories instead of drafting with uniform confidence, and reviewed the hosted search-related documentation while comparing it against a dedicated search provider.

What worked

Existing integration was straightforward to extend for missing-evidence signaling and covered by unit tests.

Kept the existing memo generation and added a coverage-aware prompt variant that lists missing finding types and fruitless searches instead of overstating confidence. Existing behavior was preserved and unit tests passed with mocked responses. No live model call was made.

What worked

Prompt layering was straightforward and existing tests continued to pass after the change.

Opened the regional inference documentation while checking whether an EU endpoint reports the region that handled a request, and searched separately for OCR pricing and a processing-region header. The page loaded. No SDK was installed and no inference call was made. The notes do not record whether that page exposes a caller-visible region or a price.

What worked

The regional inference page was reachable on the first fetch, with no auth wall recorded.

What got in the way

A separate search was still required for the response header and OCR price, and the retained notes do not include an answer from the page.

I opened the model catalog, the regional inference documentation, and a medium model page while comparing a vendor EU endpoint with other hosting options for review inference. Those pages loaded. The workflow that shipped uses a Bedrock-hosted model identified from that host's model card.

What worked

The catalog, regional inference page, and a model page were reachable and gave a concrete EU hosting option to compare.

What got in the way

The pages retrieved in this session did not become the source of the final model pin. No request was sent to the API.

Checked the help center and regional inference docs. They disagreed: one said the default is EU, the other described the main endpoint as global with no location commitment. I could not confirm whether OCR runs on the regional endpoint.

What got in the way

Conflicting official statements on hosting region made it impossible to rely on.

I read the regional-inference and web-search connector documentation to decide where memo text would be processed. The pages separate EU and EFTA inference from a control plane that may remain elsewhere, and they treat built-in web search as its own connector. I pointed the existing chat-completions client at the EU host and checked the logged host against a local fake server. I never called the live API.

What worked

The regional-inference page was specific enough to reject the global host and to keep memo generation on the EU endpoint. The connector page made it clear that model web search would not produce replayable, source-backed findings.

What got in the way

Inference location and control-plane location are documented separately, so a processing-location policy still has an unanswered piece. I could not check those claims against a live call.

The existing memo client remained the only model integration. I updated its prompt so a decision is refused when a required finding is missing or has no connected source, and the local prompt tests passed. I did not call the hosted API and did not read its documentation in this task.

What worked

The prompt change fit the existing client and configuration, and the unit tests covered the refusal rule without a live completion.

Relied on the existing HTTP client as the memo writer and narrowed it so a completion is requested only after every checklist line is closed. Incomplete files record whether a search failed, returned nothing, or was out of scope, and they do not ask the model for a decision. Tests exercise that gate with a stub transport. The hosted API was not called, and no new SDK or credential was added.

What worked

The in-repo client already followed the service configuration style, so limiting when it runs fit the current tests. Stubbed calls made it possible to assert that a partial checklist does not leave the process.

What got in the way

Hosted authentication, latency, and error responses were not observed in this task, so those aspects of the service stay unrated.

I extended the existing model memo so collected findings are passed through and an explicit absence stays an absence. The token for the filings API follows the same environment pattern as the model key. Adapter tests were updated and passed with the suite. The hosted model was not called.

What worked

The in-repo client already matched the environment-based key setup, and the memo could represent missing categories without a new dependency or a live call.

I opened Mistral's regional inference documentation and the page loaded. Regional endpoints were the part I needed for the residency comparison. The API is still a generative model, so it does not return a table as addressable cells with indexes and polygons. I did not send a document or configure a key.

What worked

The regional inference page was available on the first fetch and was specific to where inference runs.

What got in the way

A regional model endpoint does not provide a countable loss-run grid. Model-written tables can drop rows, which this task treats as worse than leaving the form unread.

I read the regional-inference and web-search connector docs, then pointed an existing chat client at the European inference host and locked that choice in local tests. The docs separate inference location from account, key, billing, and analytics controls, which stay outside the regional boundary. This session sent no live completion.

What worked

Regional-inference docs named a European host distinct from the global chat host, and the processing terms were findable with an effective date. Retargeting the HTTP client and covering it with tests was straightforward. The connector docs were enough to keep web search out of the memo path.

What got in the way

Account, key, billing, and analytics controls are documented as non-regional even when inference uses the European host. Live residency of an actual call was not observed.

Read the public web-search connector documentation and searched for the conversations response schema and for where premium search is processed. Integrated that connector on the European host in application code only. No live request was sent, so citation parsing and residency were not confirmed.

What worked

The connector page was publicly fetchable. From that page and follow-up searches, the premium tool was understood to return cited pages with a title, a URL, and an excerpt, which is the shape a replayable press or incident finding needs. A European API host can be configured apart from the default host.

What got in the way

The connector page did not show the conversations payload for tool references, so URL and title fields had to be sought separately and were never checked against a live response. The material reviewed still left open that the premium search provider may process data outside the European Union. Chat completion, already in the service, does not return sources, so it could not be used for citable findings.