Hybrid retrieval
Search live systems and the system of record, then act
Query-time Slack and CRM retrieval sits beside indexed docs, then the same turn can cite, write back, offer an assessment, or surface a process. Connector count is not the story.
Live Real Time Search
Slack, HubSpot, Jira, Salesforce, Monday, ClickUp
Indexed RAG
Docs, wikis, video, courses
Actions
Write back in the same turn
Governed answers
SemQA, assessments, ACLs
Query time, not overnight sync
Live retrieval from the tools people work in
Real Time Search runs at question time against the source, with the signed-in user's OAuth. It does not copy Slack or CRM objects into Materials. When the live hit is strong, the planner can skip the knowledge base instead of mixing stale chunks with a live record.
- Slack search via assistant.search.context, always-on when connected
- HubSpot CRM via remote MCP, ACL-aware, with record lookup
- Jira, Salesforce, Monday.com, and ClickUp at query time
- Reads stay on the Real Time Search path; write actions stay separate
Live and indexed in one answer
Reconcile the system of record with what changed this morning
Indexed RAG covers docs, wikis, video, and courses. Real Time Search covers the live systems. One agent turn can use both: SemQA and knowledge-base search for the corpus, then live connectors when the question is about a deal, ticket, or thread. Citations keep source timestamps so newer evidence is visible.
- Tool-loop RAG: SemQA → intent → rewrite → knowledge-base search (GraphRAG Cypher in parallel with vectors)
- Real Time Search planner orders connectors by relevance and can skip KB when live is enough
- Combined citation bands from KB hits and each live connector
- Per-user OAuth; query-time only — not a second copy of Slack or CRM
Do, not only find
Grounded actions beside retrieval
Chat tools execute in the same turn as retrieval. Calendly, Gmail, Zoom, Google Calendar, and write paths for Jira, Salesforce, Monday, and ClickUp run as actions. Project and personal sessions can coexist. Widget and authenticated scopes are allow-listed separately.
- Action toolkits with project shared + user/widget personal sessions
- Allow/deny policy per project, distinct auth vs widget scopes
- Real Time Search-owned reads stay off the action path so search and write do not collide
- Assessment offer tools can hand a learner a scored check from chat
Structure the corpus, then mine how work runs
Knowledge graph on ingest, GraphRAG at question time
New projects extract entities and relationships into Neo4j by default. That graph is editable in Materials. Separately, process discovery builds an OCEL event log and process tree from the same documents so SOPs and tickets become a candidate workflow, not only chunks. At question time, search_knowledge_base runs Cypher in parallel with Pinecone: ranking, co-occurrence, paths, and provenance facts land in the answer.
- Graph extraction on ingest, default-on for new projects
- Query-time GraphRAG: Cypher adjunct inside the chat KB tool (fail-open on deadline)
- Entity and edge edit, merge, and delete in the document graph UI
- Process discovery: events → OCEL → process tree / BPMN candidates; promoting a candidate to an executable workflow is still partial
Answers you can review
Semantic Q&A, assessments, and source-mirrored ACLs
Curated answers sit in front of RAG when a human has already approved the response. Assessments can be authored, published, and offered from chat or a widget. Retrieval is filtered by principal ACLs stamped at ingest or mirrored from the source. That is permissioning of knowledge, not a chatbot that sees every uploaded file.
- Semantic Q&A lookup before or beside knowledge-base search, with human review
- Project assessments: author, start, evaluate, complete; chat and widget offer
- Blind-index ACL filters on retrieval; source-mirrored grants where the provider supports them
- Citation viewer with source timestamps for KB, Slack, and HubSpot hits
Enterprise search and Work AI
How hybrid retrieval compares to enterprise search tools
These platforms win on connector catalogs and, in Glean's case, agents. The gaps are query-time live retrieval mixed with an index, process mining, governed Q&A, and assessments. Connector count is not a row here — Glean and GoSearch would win it.
Comparison as of August 17, 2026
| Capability | Tricky Wombat | Glean | GoSearch | Coveo | Guru | Elastic |
|---|---|---|---|---|---|---|
| Query-time live retrieval (not overnight index) | Yes | - | Yes | - | Partial | - |
| Live + indexed retrieval in one answer | Yes | Partial | Yes | - | Partial | - |
| Per-user OAuth to Slack / CRM / work trackers | Yes | Partial | Yes | Partial | - | - |
| Source-mirrored ACLs on retrieval | Yes | Yes | Yes | Yes | Partial | Yes |
| Agentic write-back / actions from chat | Yes | Yes | Yes | - | - | - |
| Knowledge graph extraction on ingest | Yes | Yes | - | Partial | - | - |
| Query-time GraphRAG in the chat loop | Yes | Partial | - | - | - | - |
| Citations with source timestamps | Yes | Partial | Partial | Partial | Partial | - |
| Human-reviewed curated answers (SemQA / Cards) | Yes | - | - | - | Yes | - |
| Assessments offered from chat or widget | Yes | - | - | - | - | - |
| Process mining / workflow discovery from docs | Yes | - | - | - | - | - |
| Cross-source contradiction arbitration | Partial | - | - | - | - | - |
- 1.Glean is an indexed Work AI platform with a knowledge graph and a mature agent library. It is not query-time federated search. Agents and connectors are parity — do not pitch those as TW differentiators against Glean.
- 2.GoSearch is the closest live-retrieval architecture (federated, source ACLs, actions). Beat it on process mining, assessments, curated SemQA, and (when shipped) contradiction arbitration — not on "we're live."
- 3.Guru added a Slack MCP integration in March 2026 and is honest by design (verified Cards). Honesty and curated answers are parity. Query-time CRM/work-tracker Real Time Search, process mining, and assessments are not.
- 4.Coveo security trimming is a genuine strength (ACL row is Yes). Live Real Time Search mixed with an index, assessments, and process mining are not Coveo product features.
- 5.Elastic is a retrieval toolkit (hybrid BM25 + dense). Recency arbitration, agents, graph chat, and process mining are DIY, so they rate No as productized features.
- 6.TW "Query-time GraphRAG" is Yes: search_knowledge_base invokes graph-api query_rag (Cypher) in parallel with Pinecone. Ranking answers use structured graph facts; document [N] citations come from graph-seeded chunks and/or Pinecone. Fail-open if Neo4j misses the deadline. TW "contradiction arbitration" is Partial: Real Time Search capability descriptors exist and are not wired into answer generation.
- 7.Website crawl, JavaScript render, and per-URL Watch are compared at /what-we-do/web-ingestion.
Legend: Yes = productized support · Partial = limited, optional, incomplete, or DIY · No = not a documented product feature. Website crawl and Watch are compared on the Web Ingestion page, not here.
Hybrid retrieval FAQ
What to pitch, and what not to
Web Ingestion is crawl, JavaScript render, and Watch freshness. This page is what happens at answer time: live retrieval, indexed RAG, actions, graph extraction, process mining, curated Q&A, and assessments. Connector catalogs belong to Glean and GoSearch; they are intentionally not a row here.
Live, at question time, with the signed-in user's OAuth. Slack, HubSpot, Jira, Salesforce, Monday.com, and ClickUp are query-time connectors. They do not ingest those objects into Materials. Docs, wikis, video, and courses still go through ingest and Watch.
We extract entities and relationships into Neo4j on ingest and you can edit that graph in Materials. Chat also runs GraphRAG: Cypher executes inside the knowledge-base tool in parallel with vector search. Structured graph facts (rankings, co-occurrence, paths) are the primary answer for those question types; document citations still come from retrieved chunks. Glean uses graph relevance in ranking. Pitch both extraction and query-time graph search.
Both. Retrieval tools and write actions run in the same agent turn. Calendar, mail, and work-tracker writes are allow-listed per project. Real Time Search keeps the read path so search and write do not share one brittle integration.
Do not claim more connectors, better agents, or (against GoSearch) live retrieval as a unique wedge. Glean wins catalog and agents. GoSearch shares the live posture. TW's shipped wedges here are hybrid live+index in one citation set, query-time GraphRAG, process mining, assessments, and human-reviewed Semantic Q&A.