The SHARE Research Portal is a publication-discovery and research-support environment for the SHARE (Survey of Health, Ageing and Retirement in Europe) community. It brings together semantic search, a browsable publication index, topic-based exploration, and SARA — the AI research assistant that works over the SHARE-related scientific literature.

Who runs it, and what it answers from

SARA is built and run by SHARE Austria and runs on SHARE's own servers. It answers from the publication index, the released microdata and the codebooks — not from the open internet.

The model is Qwen3.8-Flash-Next, running on this machine, so no outside AI service reads what you type or what it reads on your behalf. A search for papers outside SHARE's collection sends its search words to the public publication indexes, which the tool list beside the message box names. It holds up to 524,288 tokens of context, which is several long reports and the conversation around them, so a long working session does not quietly lose its earlier half. The same model does the rewrite behind the wand beside Send.

One assistant, and what your account can reach

There is one SARA. What differs is the set of tools your account carries. A researcher account can search the literature, look variables up, ask for aggregate figures and read files you attach. A SHARE staff account carries those and two more, the web and SHARE's own copy of a paper, because staff use the same assistant for correspondence and admin as for research.

There is no mode to switch and no setting to find. The tool list beside the message box shows exactly what is available in the conversation you are in.

Help us improve

Conversations are recorded anonymously and used to improve the assistant. Rating an answer with the thumbs is useful; saying what was wrong is far more useful. One line of correction is worth more to us than a hundred silent downvotes.

Getting help

Anything that looks broken rather than merely wrong belongs in the support inbox. The service status pill in the top bar shows whether the problem is yours or ours.

What it does

• Research assistant (SARA): conversational access to the publication corpus with citation support, summarisation, and methodology questions grounded in the SHARE literature — and to the SHARE microdata, for questions about variables, cross-wave harmonisation and analytical workflows. There is one SARA and no mode to choose: she has the literature tools and the data tools in the same conversation, and which further tools she can reach follows from your account rather than from anything you switch.
• Rewrite a draft (the wand beside Send): type or paste something you have already written, press the wand, and pick a style — journal article, project report, policy brief, internal note, or describe the register you want in your own words. The rewrite is done by the same model that answers you, and it comes back in the conversation so you can read it before you do anything with it. It edits wording, not facts: it does not check whether a claim is true, look anything up, or add to what you wrote — so read it against your original. Nothing leaves this service, and nothing is sent anywhere until you send it.
• Search publications: semantic and hybrid (BM25 + dense vector) search over the full text of indexed publications, with optional neural reranking.
• Browse by topic: OpenAlex-taxonomy-based topic tree (domain → field → subfield → topic) with publication timelines.
• Publication index: bibliographic inventory compiled from the SHARE repository crawler, including entries whose full text is available and those known by reference only.

Semantic search — what it is, and why to use it

Traditional keyword search only finds papers that contain the exact words you type. Semantic search goes a step further: queries and paper passages are both converted into high-dimensional numerical vectors (embeddings) that capture the meaning of the text rather than its surface form. Two passages that talk about the same concept end up close together in that vector space even if they share no vocabulary.

This has several concrete advantages for literature discovery:

• Paraphrase and synonym tolerance: a query for “cognitive decline in old age” also surfaces papers written as “neurocognitive ageing trajectories” or “memory loss in the elderly”.
• Natural-language queries: you can ask full questions (“how does early retirement affect depression risk?”) instead of guessing the right keyword boolean.
• Multilingual robustness: because the embedding model places semantically related terms close together, near-synonyms and closely related concepts across languages or disciplines tend to match.
• Better recall on conceptual questions: semantic search finds relevant material that keyword search would miss because the authors used different terminology.

The portal uses a hybrid retrieval strategy: semantic (dense-vector) scoring is combined with classical BM25 keyword scoring, so you keep the precision of exact term matches (for example specific variable names, author names, or waves) while also benefiting from meaning-based recall. An optional neural cross-encoder reranker then re-scores the top candidates by looking at the query and each passage together, which sharpens the ordering of the final results.

Data sources

• SHARE document repository (scientific publications, working papers, methodology reports, questionnaires, codebooks).
• Bibliographic inventory produced by the SHARE repository crawler (repository_inventory.bib), enriched with DOI and OpenAlex-topic assignments where available.
• Topic assignments derived from the OpenAlex topic taxonomy; papers without an OpenAlex match receive zero-shot assignments as a fallback.

How it works

• Text extraction: PDF content is extracted, cleaned, and chunked before indexing.
• Embeddings: chunks are embedded with BAAI/bge-m3 — a multilingual model, chosen because the SHARE corpus is not written only in English — and stored in a FAISS vector index.
• Hybrid retrieval: queries are answered by combining BM25 lexical scoring with dense-vector similarity; an optional cross-encoder reranker (BAAI/bge-reranker-large) refines the top results.
• Research assistant: SARA runs on Qwen3.8-Flash-Next, hosted on this server, with a 524,288-token context window. It answers with retrieval against the publication index, the released microdata and the codebooks, and cites the sources it draws from.
• Checking a draft questionnaire item: SHARE's 5,864 source-language questions are embedded with the same model, so an item being written can be matched against everything already fielded — which waves asked it, how the wording moved, and what changing it would cost in comparability.
• Topic exploration: per-paper topic assignments are aggregated into a domain/field/subfield/topic tree and visualised as a publication-per-year timeline.

Index refresh

The index is refreshed on a regular schedule. New publications added to the SHARE repository are picked up by the next indexing pass, which reconciles the docstore against the live repository, re-runs text extraction and embedding for new files, and rebuilds the topic assignments.

Access

Access to the research portal is limited to registered SHARE-ERIC data users in good standing. Use of the service is governed by the SHARE-ERIC Conditions of Use and is restricted to scientific research consistent with each user's registered SHARE project.

Technical stack

Vector store

FAISS

Embeddings

BAAI/bge-m3 — multilingual, 1024 dimensions

Lexical index

BM25

Reranker

BAAI/bge-reranker-large cross-encoder

Back end

FastAPI (Python)

Front end

Static HTML, CSS and JavaScript — no build step

Assistant model

Qwen3.8-Flash-Next, 524,288-token context

Rewrite model

Qwen3.8-Flash-Next — the same model, for the wand beside Send

Assistant inference

Locally hosted, on this server. Nothing is sent to a third party

Topic taxonomy

OpenAlex, with zero-shot fallback

Using SARA for data work

There is nothing to switch on and no mode to choose. Ask about the microdata in the same conversation you ask about the literature, and change subject mid-conversation if that is where the work goes — which it usually is, since a question about a variable and a question about the paper that used it are the same question from two sides.

What that looks like in practice:

• Finding the variable you mean from a description of it, rather than from its name — and reading the question as it was actually asked, in any of the 42 fielded languages.
• Checking comparability before you pool waves. Wording moves between waves, and SARA shows you where it moved rather than asserting that a series is safe to combine.
• Running the analysis. Python and R execute on this server against the release, and the code that produced a number is shown alongside it, so you can check the working rather than trust the answer.
• Working on your own files. Attach a spreadsheet or a document and she will read it, work on it, and hand back a file.

One limit is worth knowing before you start. Results are aggregate — counts, means, distributions, models — and never a single person's record. What you upload stays on this server. The files of a conversation, yours and the ones she makes, are in the Files panel, and you download them from there. Out of the secure environment a file leaves only through an export request.

No setup beyond being signed in. Which further tools SARA can reach follows from your account, and the tool list beside the message box states it exactly.

Connecting external AI tools (MCP)

SHARE is exposed as a Model Context Protocol server, so external assistants (Claude.ai, ChatGPT, Claude Desktop, any MCP client) can query the literature and the survey documentation mid-conversation — and get back records rather than a plausible-sounding guess.

What it answers:

search_share_variables

Which variable measures a construct — across 57,679 variable labels and the wording of the questions behind them.

get_share_question

What a question actually asked, as fielded in any of 42 language versions — including regional variants such as CH-de, BE-fr, IL-ar and LU-pt. Your model never has to translate SHARE wording, because it can ask for the original.

compare_share_waves

Whether an item was worded the same way across waves — the check that should happen before anyone pools them.

search_papers, get_topic_tree, get_topic_timeline, search_by_topic

The publication index and its OpenAlex topic taxonomy.

Metadata only. No microdata is reachable through this server — not filtered, not aggregated, not one row. That is a property of what the endpoint can execute, not a policy it promises to follow. Analysis stays inside the portal, where you are signed in and the data never leaves.

Two habits are built in. Every record states which documentation release it came from, because documentation and data are not always the same vintage. And where something is not documented, the answer says so and returns what is — constructed variables and national add-on questionnaires sit outside the central instrument, and a label is not a question.

Connection details:

MCP server URL

https://research.share-austria.at/mcp

Authentication

OAuth 2.0 (authorization code + PKCE)

OAuth client ID

your own portal Client ID — the same one you sign in with above.

OAuth client secret

your own portal client secret. If you have mislaid it, ask the portal administrator to reissue one; it is stored only as a hash and cannot be looked up.

Use your personal credentials, never a shared one. Every connection is attributable to one account, which is what makes access revocable for a single user without disrupting anyone else, and what keeps the usage metrics meaningful. Do not paste your secret into a shared machine, a group chat, or a config file you commit to a repository.

Claude.ai (web):

1. Open claude.ai/settings/connectors and choose “Add custom connector”.
2. Paste the server URL, then your own portal Client ID and client secret.
3. Choose Connect; you are redirected through the OAuth flow and back to Claude.
4. Ask, for example, “which SHARE variable measures life satisfaction, what exactly was asked in Italian, and was it worded the same way in wave 4?” — three tool calls, three records, no paraphrase.

ChatGPT: under Settings → Integrations (or Custom Actions), choose “Add MCP server”, paste the same credentials, and complete the OAuth authorisation.

Claude Desktop: add the following to your claude_desktop_config.json (~/Library/Application Support/Claude/ on macOS, ~/.config/Claude/ on Linux, %APPDATA%\Claude\ on Windows), then restart the app:

{
  "mcpServers": {
    "share-research-portal": {
      "url": "https://research.share-austria.at/mcp",
      "auth": {
        "type": "oauth2",
        "client_id": "YOUR-PORTAL-CLIENT-ID",
        "client_secret": "YOUR-PORTAL-CLIENT-SECRET"
      }
    }
  }
}

External MCP clients are subject to the same SHARE-ERIC Conditions of Use as the portal itself — only registered SHARE users in good standing may connect, and all queries are logged with the same privacy-safe usage metrics as direct portal traffic.

Using SARA's model in your own tools (OpenAI-compatible API)

The model behind SARA also answers at an OpenAI-compatible address, so a coding assistant in your editor (Zoo Code, Continue), a script that uses an OpenAI library, or any other client that speaks that protocol can use it with your portal login. The model runs on this server; your requests do not go to an outside AI provider.

Connection details:

Base URL

https://research.share-austria.at/openai

API key

your own portal client secret, the one you sign in with above (it begins psc_). The authenticator code is not needed here.

Model

sara

Context window

524,288 tokens

When your client brings its own tools, as an editor does to read and change your files, the model answers and your client carries out the steps on your computer. Nothing on your computer is reachable from this server. A request that sets "tools": true instead gets SARA's own tools, the ones your account has in the portal, working in your own files here.

The model is shared with everyone who uses SARA, so when it is busy a request waits for its turn.

Zoo Code (VS Code): in its settings choose the API provider OpenAI Compatible, and enter the base URL, the API key and the model above. Set the context window in the model's settings to 524288, or Zoo Code assumes a smaller one.

Continue (VS Code, JetBrains): add the model to your config.yaml:

models:
  - name: SARA
    provider: openai
    model: sara
    apiBase: https://research.share-austria.at/openai
    apiKey: YOUR-PORTAL-CLIENT-SECRET
    roles:
      - chat
      - edit
      - apply

Python, with the openai package:

from openai import OpenAI

client = OpenAI(
    base_url="https://research.share-austria.at/openai",
    api_key="YOUR-PORTAL-CLIENT-SECRET",
)
reply = client.chat.completions.create(
    model="sara",
    messages=[{"role": "user",
               "content": "What does a fixed-effects model control for?"}],
)
print(reply.choices[0].message.content)

The rules for MCP above apply here too. Use your own credentials, never a shared one, and keep them out of any file you commit to a repository; an editor's settings file is easily one of those. Use it for your work with SHARE, under the same SHARE-ERIC Conditions of Use as the portal. Every request is attributed to your account.

Contact

Questions, bug reports, or feedback can be directed to the SHARE Austria research support team.