Squirro connects enterprise search with knowledge graphs and AI workflows

Squirro builds enterprise search and generative-AI workflows around a structured understanding of organizational knowledge. Its platform combines semantic retrieval, taxonomies, knowledge graphs and agents, with controls for permissions and private deployment. The opportunity is to make existing information more useful across a business process. The hard part is deciding which relationships and source documents are authoritative enough for the system to rely on.

01 / ProductKnowledge structure connects search to downstream AI work

The Enterprise GenAI Platform describes layers for security, graph-informed context, search and agents, and enterprise integration. The intention is to reuse the same governed knowledge across finding information and acting on it. That is broader than adding a language-model answer above a list of search results.

The company overview explains that Squirro acquired Synaptica in 2024, bringing taxonomy and knowledge-graph management into the portfolio. The coverage identity remains Squirro. A taxonomy organizes concepts; an ontology describes their relationships; a knowledge graph connects actual entities and records using that structure. Those distinctions matter because maintaining the concepts is different work from indexing documents.

Squirro’s knowledge-graph explanation connects domain modeling, entity tagging and classification with GraphRAG. GraphRAG uses structured relationships to inform retrieval before generation. It can help express a connection that ordinary similarity search misses, but the resulting answer still depends on correct source data, correct relationships and appropriate interpretation by the model.

02 / AudienceA candidate where institutional knowledge has a recognizable structure

An enterprise with approved proposal material, technical documentation and customer-specific restrictions is a plausible audience. The same term may have different meanings across products or departments, and a useful answer may require linking a capability to its supported product version. A graph and taxonomy can represent those distinctions if someone is responsible for keeping them current.

Squirro is less suitable when no team can establish which content is approved or when the immediate need is simply searching a small folder. Adding semantic structure creates a maintenance obligation. The value comes from resolving real ambiguity and making reusable relationships explicit, not from having a graph for its own sake.

The Neo4j blueprint provides a relevant comparison for graph infrastructure and graph-backed application development. The Glean blueprint offers another approach to employee knowledge access. Compare the amount of domain modeling your use case needs and whether you want to build on infrastructure or adopt a broader enterprise application.

03 / WorkflowProposed workflow: draft RFP answers from approved reusable knowledge

A proposed pilot would help a bid team find approved material and prepare draft answers to an RFP. It should preserve a distinction between reusable company evidence and the prospective customer’s requirements. A question in an incoming tender is not proof that the supplier already has the requested capability, and the system must not turn it into such proof.

1. 01

   Define the approved answer library

   Select product documentation, current security statements and reusable case material with named owners. Record the product version, permitted audience and last approval for reuse. Exclude old proposals whose promises were specific to one customer or commercial agreement.

2. 02

   Model the domain vocabulary

   Agree on product names, capabilities, certifications and relationships to supporting evidence. Identify synonyms and similarly named features that must remain distinct. Use subject experts to review extracted relationships before they influence answer selection.

3. 03

   Extract and classify requirements

   Process an example RFP into questions, mandatory requirements and submission rules. Check the extraction against the original document. A missed mandatory requirement is an intake failure even if every generated answer reads well.

4. 04

   Retrieve evidence and draft responses

   Ask for the best approved supporting material for each question and a draft that cites it. Mark missing evidence explicitly. Keep unsupported commitments out of the answer, especially around features, delivery dates and security assertions.

5. 05

   Review, approve and retain the outcome

   Route answers to the appropriate experts and record changes. Compare review time, missing requirements and unsupported claims with the existing process. Only approved reusable improvements should return to the knowledge library.

The RFP and proposal agent page describes intake, retrieval and cited drafts with human review. Its interactive examples are vendor demonstrations, not independent evidence of customer results. The useful design idea is that an answer carries its source and approval context through review rather than becoming a detached paragraph that nobody can verify.

For this pilot, make conflicting evidence a first-class test. A current product document may disagree with a successful old proposal. The system should prefer current approved evidence or flag the conflict, not assume the historical winning bid is authoritative. Relevance to the question and permission to reuse a statement are separate properties.

The enterprise-search page describes semantic retrieval, content classification and access controls. These capabilities should be evaluated together. Finding an excellent answer in a restricted customer document is still a failure if the current user lacks permission to use it in a new proposal.

04 / PricingCommercial scope depends on the knowledge domain and deployment

The contact page invites pricing and packaging discussions based on scale and deployment needs. A useful proposal should therefore describe the actual corpus and business process, not simply name an agent. Ask which connectors, transformations and approval interfaces are included, and how the commercial scope changes when another department adds a different knowledge domain.

Squirro’s on-premises material describes running the platform within customer infrastructure, including restricted environments. Evaluate the actual architecture and model route rather than importing broad claims about complete sovereignty into your own situation. The operating plan must account for updates, support access, backups and any external services the workflow uses.

05 / DistinctionsTaxonomy management can make retrieval more deliberate

The combination of a knowledge-management foundation and generative workflows is the meaningful distinction. A bid team can define that a capability belongs to a particular product edition and requires a specific deployment. Those relationships help an assistant avoid treating every mention of the capability as equally relevant to every question.

This structured approach also creates a correction surface. If the wrong product is linked to a capability, an expert can fix the relationship or underlying source instead of repeatedly rewriting prompts. That is potentially valuable for long-lived enterprise knowledge, where stable meanings matter even as models and interfaces change.

The benefit has a cost: someone must maintain the definitions and resolve disagreements. A graph can encode an incorrect assumption as neatly as a correct one. The strongest evaluation shows how a domain owner identifies and fixes that mistake and how the corrected relationship changes subsequent search results and draft answers.

06 / QuestionsDistinguish governance features from guarantees of correctness

Squirro uses strong language about deterministic accuracy and eliminating hallucinations. This blueprint does not treat those claims as guarantees. A graph can constrain retrieval and improve context, but it does not establish that every extracted relationship or generated statement is correct. Evaluate missing evidence, contradictory sources and plausible but unsupported answers explicitly.

The security page provides an important qualification: it states ISO 27001 certification and SOC 2 Type I attestation, with Type II in progress. Do not describe Type II as complete. The same page discusses ACL filtering, SAML-based SSO, audit logs and encryption. Obtain the relevant current reports and inspect how those controls apply to the selected deployment.

Permissions also need testing across derived content. A draft answer can contain information from several documents, and its visibility should respect the original access boundaries. Confirm what happens when a source is removed, access is revoked or an employee changes role. Search-time filtering alone does not explain the lifecycle of every previously generated artifact.

07 / DecisionBegin with a knowledge domain that has an accountable owner