Spheriq AI works as a guided pipeline system and is therefore more complex than a freely improvising chatbot. This architecture is crucial if AI is to be used reliably, data-efficiently and comprehensibly in the non-profit sector. This background article explains why and how Spheriq AI works with prompt launchers and an eight-stage pipeline.
Most AI tools start with an empty input field. The users themselves must know what information is relevant, how to formulate the question, what data they are allowed to insert and how the answer must be checked at the end. This may be sufficient for general and one-off tasks. However, it is not enough for everyday fundraising, funding or application checks.
Why a chatbot is not enough
Spheriq AI works with both public and confidential institutional data. This includes, for example, project data on projects that are not yet public, applications, strategy documents and personal data. This data is only visible to certain users and roles – for good reason. An AI that does not know the data set precisely is of modest use. And manual copy-paste quickly becomes time-consuming and risky.
Above all, Spheriq is a toolbox. The benefits develop in everyday life and in the specific work context. A question on an organizational profile may suggest a different answer than one in a search, on a project or in a carefully constructed list or in the case of an advanced request. Spheriq AI must therefore not only understand the question, but also take into account the location of the query, the role of the user and thus the purpose of the respective work step.
Precise launchers instead of mumbo jumbo
That’s why Spheriq AI often starts in everyday life via so-called launchers. These are predefined entry points directly in the respective work context, for example for matching, profile enhancement or funding research. This way, users don’t have to learn how to write a well-functioning AI prompt first. They select a suitable action and Spheriq AI takes the current context with it.
A launcher is more than just a button. It contains a technical logic: What should be checked? Which data may be used? Is it about your own profile, an external organization, a project, funding criteria or a search? Spheriq AI does not start from scratch, but with a guided order.
The eight stages of the pipeline
Clicking on the launcher or sending an individual prompt starts the Spheriq AI pipeline. It is divided into eight separate steps, some of which are broken down again into individual sub-steps. Each step has its own task and reduces a different risk: incorrect context, incorrect data access, incomplete evidence, generic answers or incomprehensible conclusions.
1. planning
The pipeline first classifies the purpose of the request. It distinguishes, for example, whether it is a self-analysis, a matching, a funding search or a document analysis. At the same time, it separates the starting point and goal of the query (“scope” and “target”). This sounds simpler than it is in practice because the search direction is highly context-dependent. Should your own profile be evaluated, compared with someone else’s or should the other profile be checked for fit? Careful “request analysis” ensures that the right course is set. It avoids a profile analysis being seen as a matching process. This paves the way for the next steps.
2. Werkzeug-Einsatz vorbereiten
Auf Basis des Plans werden später konkrete Werkzeuge eingesetzt. Diese Werkzeuge brauchen aber klare Anweisungen: Welche Organisationen sind in organization_search konkret gesucht? Geht es um Themen, Zielgruppen oder Wirkungsgebiete? Kurz: Die Werkzeugparameter müssen in der «Tool Argument Resolution» vorbereitet werden. Dazu gehören Entitätskennungen, Suchfilter, Kategorien oder Evidenzpräferenzen. This step is particularly important for searches: If a nonprofit organization is looking for suitable funding organizations, the search type, topics, target groups and impact area must be correctly derived and coded.
3. retrieve tools
The required tools are now used. The pipeline therefore independently retrieves profiles, search results, list contents or documents, in several sub-steps in the case of extensive databases. The returned data, documents and search results are then converted into a structured form so that they can be properly interpreted and processed in the subsequent steps. This sub-step is called “Collection Item Preparation”.
4. select candidates
Now the table is set – but not every result has the same significance for further processing. For larger result sets, the pipeline filters and prioritizes the retrieved elements in a next step. Weak or irrelevant hits are removed so that they do not affect further evaluation. This so-called “candidate selection” is based on the starting point and objective of the query: For example, different signals count for a funding search than for a profile check or a document summary.
5. check intention
The next important control step is “target selection”: the pipeline checks again whether the correct target entity or target collection has been found. It also distinguishes exact hits from partial hits and checks whether expected hits are completely missing. It therefore clarifies whether the starting point and target have been resolved correctly. This step prevents an answer from being based on an incorrect or only seemingly appropriate foundation.
6. condensing the basics
The amount of data prepared at this stage can still be very extensive: Dozens of search results and hundreds of thousands of characters in loaded documents are ready for evaluation. This is why the “Batch Reducer” is now used to condense larger result sets. As this is a delicate intermediate step, a subsequent work step, the so-called “Exact Collection Rewrite”, ensures that the relevant evidence is retained and the counting method is correct. As a result, the basis for the answer is now available in compact form. This ensures that the AI is not simply “swimming” in a huge amount of data, but can work with an organized evidence base.
7. Evaluation
Importantly, this is where the actual expert assessment begins. The pipeline now interprets the evidence and, in light of the original question, evaluates what is “strong”, what is “weak”, what genuinely fits together, and which conclusions can be drawn from the available information. This is where a substantial amount of philanthropy-specific expertise comes into play: funding logics, fit, hope, badges, profile quality, matching, preliminary application reviews, or funding criteria are not assessed generically, but according to a carefully designed logic (see Part 3 of the background series: Keys to AI).
At this stage, the “Final Answer Builder” produces an initial version of the response. However, this version is usually more detailed than necessary and, above all, has not yet been consistently condensed into the desired output format.
8. answers
Finally, the answer is cleaned up in the “Final Answer Rewrite” and adapted to the desired format. The language and output format are checked, internal details are removed and inconsistencies are corrected (final answer repair). This is crucial, as different answer logics apply to different use cases. A profile analysis, for example, requires a different structure than a matching, a badge assessment or a summary of eligibility criteria. The pipeline therefore not only provides an answer, but also a suitable format. The anchoring in the available evidence is also checked again here (verifier). If necessary, the pipeline goes back to previous steps. The final answer should be understandable, concise, useful and comprehensible: without AI slang, without exaggeration and without flattery (so-called “sycophancy”).
The Spheriq AI pipeline is not an end in itself. It fulfills several tasks simultaneously: it controls data access, reduces hallucinations, strengthens traceability and ensures that philanthropy-specific concepts are applied consistently. It also helps to uncover missing, contradictory or inaccessible information in good time and to name it clearly instead of filling gaps with assumptions. The pipeline is therefore more than just quick text production.
Spheriq AI thus follows the principles that are currently being discussed under terms such as “Responsible AI” or “Explainable AI”: AI should be supportive, limit access to data, make answers comprehensible and not replace human responsibility. This is precisely where the EU Commission’s guidelines for trustworthy AI come in: With human oversight, technical robustness, data protection, transparency and accountability.
What users get out of this architecture
For users, the pipeline means above all: less prompt work, more precise contexts and more specific and comprehensible answers. If Spheriq AI uses documents in addition to organizational or project profiles, these are listed and summarized at the end of the response. This makes it clear what an assessment is based on.
In short: anyone using Spheriq AI should not have to think about which data can be copied or how a technically precise question is structured. The launcher and the pipeline take the lead, but usage remains flexible. After an answer, work can continue in the chat and the results can be deepened, shortened or further processed.
The pipeline does make the AI slower than some general chatbots. At the same time, it makes it more reliable, more precise and more stable. And therein lies its purpose. The result is not a quickly improvised AI response, but guided, verifiable and context-related support for day-to-day work in the non-profit sector.
