Knowledge Graph Query Lab

Lightweight LLM call (Groq / Llama 3.3 70B, temperature 0.1) classifies the query into one of four intents. Returns intent + confidence + reasoning. ~200ms latency.

Intents:
  fund   — investors, raising money, LPs, VCs, funding rounds
  hire   — recruiting, talent, skills, engineers
  sell   — finding customers, leads, clients, sales targets
  general — everything else (falls back to baseline behavior)

Traverses 7 edge types: INVESTED_IN, LEAD_INVESTED_IN, FOLLOW_ON_INVESTED_IN, INVESTMENT_PARTNER_IN, INVESTMENT_PARTNER_AT, BOARD_MEMBER_OF, RAISED

Returns: seed entities + typed investment connections with names, UUIDs, labels.

Traverses 4 edge types: WORKS_AT, HAS_EXPERTISE, FOUNDED, ADVISOR_TO / BOARD_MEMBER_OF

Returns: people with their skills, companies, founded orgs, advisory roles.

Traverses 4 edge types: WORKS_AT, HAS_EXPERTISE, LOCATED_IN, IN_REGION

Returns: companies/people with industry, expertise, location connections.

Traverses all edges. Same behavior as existing semantic-question function. This is the fallback — ensures we never do worse than the current system.

Full pipeline in one POST. Body: {"task":"Find me Chief Technology Officers at AI startups"}. Returns trace_id, tool, params, cluster (pre-filter metadata), weights (per-tool ranker weights), ranked_rows (each with composite_score + rank_breakdown), synthesized result, and latency_ms.

# 1. Run the agent
curl -X POST "https://xh2o-yths-38lt.n7c.xano.io/api:LITebdJ-/robert-lab/agent" \
  -H "Content-Type: application/json" \
  -d '{"task":"Find me Chief Technology Officers at AI startups"}'

# 2. Capture trace_id from response, then give feedback
curl -X POST "https://xh2o-yths-38lt.n7c.xano.io/api:LITebdJ-/robert-lab/feedback" \
  -H "Content-Type: application/json" \
  -d '{"trace_id":"1776374223250-0d033dca9c","signal":"up","note":"Adar was spot-on"}'

# 3. Read back the trace + all feedback
curl "https://xh2o-yths-38lt.n7c.xano.io/api:LITebdJ-/robert-lab/trace?trace_id=1776374223250-0d033dca9c"

Live example result — 5 CTO candidates ranked by composite: Adar Arnon 0.945 → Adam Brown 0.855 → John Fitzpatrick 0.855 → Dharmesh Shah 0.765 → Eric Iverson 0.735. Synth LLM orders output by composite desc; confidence tiers (high/medium/low) track composite.

B consistently returned more results (top-20 vs top-10 vector search) while keeping them focused on the intent. More results, less noise.

For fund-intent queries, B surfaces INVESTED_IN, BOARD_MEMBER_OF, RAISED, and FOLLOW_ON edges that A drowns in HAS_EXPERTISE noise. The signal-to-noise ratio jumps from 4% to 75%.

The baseline fallback works perfectly. When the classifier detects a general query, B produces the exact same results as A. We never do worse.

Query #6 ("ML people at startups") returned 87 edges from A, 84 of which were HAS_EXPERTISE. The baseline blindly traverses all edges on all matched nodes, which floods the response with irrelevant connections. B avoids this by only traversing intent-relevant edge types.

115 out of 146 score≥2 investors had graph connections. Hit rate held steady from the initial 30-investor test (83%) to the full batch (78%). The knowledge graph has meaningful coverage of the medtech investment ecosystem — not just the SaaS/tech bubble.

27 matches scored below 0.4 (near-exact). 9 of the top 11 are exact company matches: OrbiMed, Earlybird, Claret Capital, Bank of America, J&J, L.E.K. These aren't fuzzy guesses — the graph reveals actual entities with real investment and team connections.

Arboretum Ventures → Arbor Ventures, Opal Venture Partners → Owl Ventures. These are semantically similar but not the same entity. A confidence threshold or exact-match flag would help separate "confirmed in graph" from "similar entity found."

The full batch scan is complete. Next step: expand to the 330 lower-confidence score-1 investors, or deploy the intent-aware query pipeline to production.

The existing production semantic-question function is running on cached results. It does not wrap vector queries in vecf32(), which means once the FalkorDB cache expires, all queries fail with "Invalid arguments". This was discovered during A/B testing — the baseline was returning stale cached results, not live queries. This is a live production bug that needs to be fixed in the existing function, independent of the intent-aware work.

FalkorDB's db.idx.vector.queryNodes requires vecf32([...]) format, not raw arrays. The existing semantic-question function appeared to work without it due to caching. Once cache expired, all queries failed with "Invalid arguments". Fix: wrap all vector queries in vecf32().

The sequence WHERE score < 0.85 ORDER BY score ASC causes a parse error — FalkorDB reads ORD as a broken OR operator. Fix: use the collect() / UNWIND pattern to separate WHERE from ORDER BY with intermediate WITH clauses.

vectors/create-vectors-string returns a JavaScript string from api.lambda, but Xano auto-parses it into a JSON array (1536 floats). The template engine {{$var.embeddings}} renders it as [0.005,-0.024,...]. Must wrap in vecf32() to make FalkorDB accept it.

Almost all nodes (11,948/13,456) share the label Entity. The real typing is in display_label field: Person, Company, VC_Firm, DomainExpertise, SubDomainExpertise, etc. This means label-based Cypher filtering won't work — must use display_label property or edge types.

Groq + Llama 3.3 70B at temperature 0.1 classifies queries in ~200ms with high accuracy. All test queries classified correctly on first try. The confidence scores are useful for fallback logic (low confidence → use general intent).