CAR-T Intelligence Agent — Architecture Design Document

Author: Adam Jones Date: February 2026 Version: 1.2.0 License: Apache 2.0

---

1. Executive Summary

The CAR-T Intelligence Agent extends the HCLS AI Factory platform to support cross-functional intelligence across the CAR-T cell therapy development lifecycle. The agent breaks down data silos between the stages of CAR-T development:

1. Target Identification — Antigen biology, expression profiling, disease association
2. CAR Design — scFv selection, costimulatory domains, signaling architecture
3. Vector Engineering — Transduction, viral vector production, manufacturing processes
4. In Vitro / In Vivo Testing — Cytotoxicity, cytokine assays, animal models, persistence
5. Clinical Development — Trial design, response rates, toxicity management

The platform enables cross-functional queries like "Why do CD19 CAR-T therapies fail in relapsed B-ALL?" that simultaneously search published literature, clinical trials, CAR construct data, assay results, and manufacturing records — returning grounded answers with clickable PubMed and ClinicalTrials.gov citations.

Comparative Analysis Mode auto-detects "X vs Y" queries (e.g., "Compare 4-1BB vs CD28 costimulatory domains"), runs dual retrievals with per-entity filtering, and produces structured side-by-side analysis with markdown tables.

Key Results

Metric	Value
Total vectors indexed	6,266 across 11 Milvus collections (10 owned + 1 read-only)
Multi-collection search latency	12-16 ms (11 collections, top-5 each, cached)
Comparative dual retrieval	~365 ms (2 × 11 collections, entity-filtered)
Full RAG query (search + Claude)	~24 sec end-to-end
Comparative RAG query (dual search + Claude)	~30 sec end-to-end
Cosine similarity scores	0.74 - 0.90 on demo queries
Manufacturing success rate (seed script)	100% (all collections populated, 0 ingest errors)

---

2. Architecture Overview

2.1 Mapping to VAST AI OS

VAST AI OS Component	CAR-T Agent Role
DataStore	Raw files: PubMed XML, ClinicalTrials.gov JSON, seed data JSON
DataEngine	Event-driven ingest pipelines (fetch → parse → embed → store)
DataBase	11 Milvus collections (10 owned + 1 read-only) + knowledge graph (25 targets, 8 toxicities, 10 mfg)
InsightEngine	BGE-small embedding + multi-collection RAG + query expansion
AgentEngine	CARTRAGEngine (retrieve → augment → generate) + Streamlit UI

2.3 System Diagram

                        ┌─────────────────────────────┐
                        │   Streamlit Chat UI (8521)   │
                        │   Cross-functional queries   │
                        │   + Comparative Analysis UI  │
                        └──────────────┬──────────────┘
                                       │
                        ┌──────────────▼──────────────┐
                        │     CARTRAGEngine            │
                        │  retrieve → augment → gen    │
                        │  + comparative detection     │
                        └──────────────┬──────────────┘
                                       │
                  ┌────────────── "X vs Y"? ──────────────┐
                  │ YES                                NO  │
                  ▼                                        ▼
        ┌──────────────────┐                   ┌──────────────────┐
        │ Comparative Mode │                   │ Standard Mode    │
        │ Parse 2 entities │                   │ Single retrieve  │
        │ Entity resolution│                   │                  │
        │ Dual retrieval   │                   │                  │
        └────────┬─────────┘                   └────────┬─────────┘
                 │                                      │
                 └──────────────────┬───────────────────┘
                                   │
                 ┌─────────────────┼─────────────────────┐
                 │                 │                      │
        ┌────────▼────────┐  ┌────▼───────────┐  ┌──────▼──────────┐
        │  Query Expansion │  │ Knowledge Graph │  │ Claude Sonnet   │
        │  169 keywords    │  │ 25 targets      │  │ 4.6 (Anthropic) │
        │  → 1,496 terms   │  │ 8 toxicities    │  │ Streaming RAG   │
        │  12 categories   │  │ 10 manufacturing │  │ + Comparative   │
        │                  │  │ 39+ entity alias │  │   prompt builder│
        └────────┬────────┘  └────────┬────────┘  └─────────────────┘
                 │                     │
        ┌────────▼─────────────────────▼────────┐
        │        Multi-Collection RAG Engine     │
        │   Parallel search across 11 collections│
        │   Weighted: lit 0.30 | trial 0.25 |    │
        │   construct 0.20 | assay 0.15 |        │
        │   manufacturing 0.10                   │
        └───┬────┬─────┬─────┬──────┬───────────┘
            │    │     │     │      │
    ┌───────▼┐ ┌▼────┐┌▼────┐┌▼───┐┌▼────────┐
    │ cart_  │ │cart_ ││cart_││cart_││ cart_    │
    │ litera-│ │trial-││cons-││assa││ manufac- │
    │ ture   │ │s     ││truc-││ys  ││ turing   │
    │ 5,047  │ │ 973  ││ts 6 ││ 45 ││   30     │
    └────────┘ └─────┘└─────┘└────┘└──────────┘
         ▲        ▲       ▲      ▲       ▲
    ┌────┴────┐ ┌─┴──┐ ┌──┴──┐┌─┴──┐ ┌──┴───┐
    │ PubMed  │ │ CT │ │ FDA │ │Pub │ │ Pub  │
    │E-utils  │ │.gov│ │seed │ │lit │ │ lit  │
    │ API     │ │API │ │data │ │seed│ │ seed │
    └─────────┘ └────┘ └─────┘└────┘ └──────┘

---

3. Data Collections — Actual State

All 11 collections (10 owned + 1 read-only) are populated and searchable.

3.1 cart_literature — 5,047 records

Attribute	Value
Source	PubMed via NCBI E-utilities (esearch + efetch)
Ingest time	~15 min
Fields	PMID, title, text_chunk, source_type, year, cart_stage, target_antigen, disease, keywords, journal
Embedding	FLOAT_VECTOR(384), BGE-small-en-v1.5
Index	IVF_FLAT, COSINE, nlist=1024, nprobe=16
Stage classification	Automated: target_id, car_design, vector_eng, testing, clinical
Target extraction	25 antigens detected: CD19, BCMA, CD22, CD20, CD30, HER2, GPC3, etc.

3.2 cart_trials — 973 records

Attribute	Value
Source	ClinicalTrials.gov REST API v2
Ingest time	~3 min
Fields	NCT ID, title, text_summary, phase, status, sponsor, target_antigen, car_generation, costimulatory, disease, enrollment, start_year, outcome_summary
Phase distribution	Early Phase 1 through Phase 3
Status	Recruiting, completed, terminated, withdrawn, active
Antigen extraction	Automated from trial title and description

3.3 cart_constructs — 6 records

Attribute	Value
Source	FDA-approved CAR-T product labels
Products	Kymriah (tisagenlecleucel), Yescarta (axicabtagene ciloleucel), Tecartus (brexucabtagene autoleucel), Breyanzi (lisocabtagene maraleucel), Abecma (idecabtagene vicleucel), Carvykti (ciltacabtagene autoleucel)
Fields	name, text_summary, target_antigen, scfv_origin, costimulatory_domain, signaling_domain, generation, hinge_tm, vector_type, fda_status, known_toxicities
Construct IDs	fda-tisagenlecleucel, fda-axicabtagene-ciloleucel, fda-brexucabtagene-autoleucel, fda-lisocabtagene-maraleucel, fda-idecabtagene-vicleucel, fda-ciltacabtagene-autoleucel

3.4 cart_assays — 45 records

Attribute	Value
Source	Curated from landmark publications
Papers	ELIANA (NEJM 2018), ZUMA-1 (NEJM 2017), ZUMA-2 (NEJM 2020), TRANSFORM (Lancet 2022), KarMMa (NEJM 2021), CARTITUDE-1 (Lancet 2021), plus preclinical studies
Assay types	Cytotoxicity (12), in vivo/clinical (9), persistence (5), flow cytometry (5), exhaustion (3), cytokine (3), proliferation (3)
Coverage	All 6 FDA products + CD22, dual-target, GPC3, HER2, Mesothelin
Resistance data	CD19 loss/mutation, trogocytosis, lineage switch, BCMA biallelic loss, sBCMA decoy
Linked constructs	Records reference FDA construct IDs where applicable

3.5 cart_manufacturing — 30 records

Attribute	Value
Source	Curated from published manufacturing data and FDA guidance
Process steps	Transduction (6), expansion (7), harvest (2), formulation (3), cryopreservation (2), release testing (6), emerging platforms (4)
Vector types	Lentiviral, gamma-retroviral, transposon (Sleeping Beauty/piggyBac), mRNA
Coverage	VCN, titer, transduction efficiency, IL-2 vs IL-7/IL-15 expansion, rapid 6-day (Kite), defined CD4:CD8 (Breyanzi), POC manufacturing, allogeneic/off-the-shelf, cost analysis, GMP facility requirements
Key parameters	Functional titer, VCN, transduction efficiency, fold expansion, viability, sterility, CAR expression, RCL, identity

3.6 Index Configuration (all collections)

Parameter	Value
Index type	IVF_FLAT
Metric	COSINE
nlist	1024 (literature), 256 (trials), 128 (constructs, assays, manufacturing)
nprobe	16
Embedding dim	384 (BGE-small-en-v1.5)

---

4. Knowledge Graph

4.1 Target Antigens (25 entries)

Each entry includes: protein name, UniProt ID, expression pattern, associated diseases, FDA-approved products, key clinical trials, known resistance mechanisms, toxicity profile, and normal tissue expression.

Target	Diseases	Approved Products
CD19	B-ALL, DLBCL, FL, MCL, CLL	Kymriah, Yescarta, Tecartus, Breyanzi
BCMA	Multiple Myeloma	Abecma, Carvykti
CD22	B-ALL (CD19-neg relapse)	—
CD20	NHL, CLL	—
CD30	Hodgkin lymphoma	—
CD33	AML	—
CD38	Multiple Myeloma	—
CD123	AML, BPDCN	—
GD2	Neuroblastoma	—
HER2	Breast, gastric (solid tumor)	—
GPC3	Hepatocellular carcinoma	—
EGFR	Glioblastoma, NSCLC	—
Claudin18.2	Gastric, pancreatic	—
Mesothelin	Mesothelioma, ovarian, pancreatic	—
+ 11 more	Various	—

4.2 Toxicity Profiles (8 entries)

Toxicity	Mechanism	Management
CRS	IL-6/IFN-γ cytokine storm	Tocilizumab, corticosteroids
ICANS	CNS endothelial activation	Dexamethasone, supportive
B-cell aplasia	On-target CD19/CD22 depletion	IVIG replacement
HLH/MAS	Macrophage hyperactivation	Anakinra, etoposide
Cytopenias	Marrow suppression	G-CSF, transfusions
TLS	Rapid tumor lysis	Rasburicase, hydration
GvHD	Allogeneic only, donor T-cells	Steroids, ruxolitinib
On-target/off-tumor	Normal tissue expression	Affinity tuning, safety switches

4.3 Manufacturing Processes (10 entries)

Lentiviral transduction, retroviral transduction, T-cell expansion, leukapheresis, cryopreservation, release testing, vector production, quality control, formulation, potency testing.

4.4 Entity Aliases (39+ entries)

For Comparative Analysis Mode, the knowledge graph includes entity aliases that resolve product names, costimulatory domains, vector types, biomarkers, and regulatory terms to canonical entities with associated target antigens.

Alias Category	Count	Examples
FDA Products	12	Kymriah → CD19, Carvykti → BCMA, tisagenlecleucel → CD19, etc.
Costimulatory Domains	4	4-1BB (CD137), CD28, 4-1BB/CD28, OX40
Vector Types	2	Lentiviral, Retroviral
Biomarker Terms	8	Ferritin, CRP, IL-6, sIL-2R, LDH, etc.
Regulatory Terms	6	BLA, RMAT, accelerated approval, breakthrough therapy, etc.
Safety Terms	7	REMS, FAERS, black box warning, post-marketing, etc.

Resolution priority: CART_TARGETS (25) → ENTITY_ALIASES (39+) → CART_TOXICITIES (8) → CART_MANUFACTURING (10)

4.5 API Functions

get_target_context("CD19")              # Returns full CD19 knowledge block
get_toxicity_context("CRS")             # Returns CRS management details
get_manufacturing_context(...)           # Returns manufacturing process details
get_all_context_for_query(text)          # Auto-detects entities and returns all relevant context
get_knowledge_stats()                    # Returns counts: {target_antigens: 25, ...}
resolve_comparison_entity("Kymriah")     # → {"type": "product", "canonical": "Kymriah (tisagenlecleucel)", "target": "CD19"}
get_comparison_context(entity_a, entity_b)  # Side-by-side knowledge graph context for dual entities

---

5. Query Expansion

Twelve expansion map categories (6 original + 6 added for expanded collections):

Category	Keywords	Expanded Terms	Examples
Target Antigen	26	196	CD19 → [CD19, B-ALL, DLBCL, tisagenlecleucel, axicabtagene, ...]
Disease	16	143	multiple myeloma → [MM, plasma cell neoplasm, RRMM, ...]
Toxicity	14	136	CRS → [cytokine release syndrome, cytokine storm, tocilizumab, IL-6, ...]
Manufacturing	16	181	transduction → [lentiviral, retroviral, VCN, MOI, viral vector, ...]
Mechanism	19	224	resistance → [antigen loss, lineage switch, trogocytosis, exhaustion, ...]
Construct	20	206	bispecific → [dual-targeting, tandem, bicistronic, CD19/CD22, ...]
Safety	15	135	REMS → [risk evaluation, mitigation strategy, FAERS, adverse event, ...]
Biomarker	14	125	CRS prediction → [ferritin, CRP, IL-6, sIL-2R, predictive biomarker, ...]
Regulatory	12	108	BLA → [biologics license application, accelerated approval, RMAT, ...]
Sequence	10	95	scFv → [single-chain variable fragment, VH, VL, CDR, binding affinity, ...]
Real-World	12	110	registry → [real-world evidence, CIBMTR, post-marketing, outcomes, ...]
Immunogenicity	10	92	ADA → [anti-drug antibody, immunogenicity, neutralizing antibody, ...]
Total	169	1,496

The expand_query() function detects keywords in the user's query and returns relevant expansion terms, which are used to run additional filtered searches across collections.

---

6. Multi-Collection RAG Engine

6.1 Search Flow (measured on DGX Spark)

User Query: "Why do CD19 CAR-T therapies fail in relapsed B-ALL?"
    │
    ├── 1. Embed query with BGE asymmetric prefix               [< 5 ms]
    │      "Represent this sentence for searching relevant passages: ..."
    │
    ├── 2. Parallel search across 11 collections (top-5 each)   [12-16 ms]
    │   ├── cart_literature:     CD19 CAR-T failure papers       (score: 0.82-0.90)
    │   ├── cart_trials:         Terminated CD19 B-ALL trials    (score: 0.74-0.85)
    │   ├── cart_constructs:     CD19 CAR designs                (score: 0.78-0.86)
    │   ├── cart_assays:         Resistance/failure assay data   (score: 0.76-0.85)
    │   └── cart_manufacturing:  Production failure modes        (score: 0.71-0.79)
    │
    ├── 3. Query expansion: "CD19" → [CD19, B-ALL, DLBCL,       [< 1 ms]
    │      tisagenlecleucel, axicabtagene, ...]
    │
    ├── 4. Expanded filtered search (top-3 per term, top-5       [8-12 ms]
    │      expansion terms, collections with target_antigen)
    │
    ├── 5. Merge + deduplicate + rank (cap at 30 results)        [< 1 ms]
    │      Weights: lit 0.30, trial 0.25, construct 0.20,
    │               assay 0.15, manufacturing 0.10
    │
    ├── 6. Knowledge graph augmentation:                         [< 1 ms]
    │      CD19 → known_resistance: [CD19 loss, lineage switch, trogocytosis]
    │      CD19 → toxicity_profile: {CRS: 30-90%, ICANS: 20-65%}
    │      CD19 → approved_products: [Kymriah, Yescarta, Tecartus, Breyanzi]
    │
    ├── 7. Build prompt: evidence grouped by collection +         [< 1 ms]
    │      knowledge context + question + citation instructions
    │
    └── 8. Stream Claude Sonnet 4.6 response                    [~22-24 sec]
           Grounded cross-functional answer with clickable
           PubMed and ClinicalTrials.gov citation links

Total: ~24 sec (dominated by LLM generation; retrieval is ~25 ms)

6.2 Collection Weights

Collection	Weight	Rationale
cart_literature	0.30	Published evidence is the primary source of truth
cart_trials	0.25	Clinical outcomes provide direct translational answers
cart_constructs	0.20	Design data explains mechanisms and structure-function
cart_assays	0.15	Lab data supports mechanistic claims with quantitative evidence
cart_manufacturing	0.10	Manufacturing links to clinical outcomes and feasibility

6.3 System Prompt

The agent uses a specialized system prompt instructing Claude to: 1. Cite evidence with clickable links — Literature citations link to PubMed ([Literature:PMID](https://pubmed.ncbi.nlm.nih.gov/PMID/)), trial citations link to ClinicalTrials.gov ([Trial:NCT...](https://clinicaltrials.gov/study/NCT...)) 2. Think cross-functionally — connect insights across development stages 3. Highlight failure modes and resistance mechanisms when relevant 4. Be specific — cite trial names (ELIANA, ZUMA-1), products (Kymriah, Yescarta), and quantitative data 5. Acknowledge uncertainty — distinguish established facts from emerging data 6. Suggest optimization strategies based on historical data

6.4 Embedding Strategy

BGE-small-en-v1.5 uses asymmetric encoding — queries and documents are embedded differently:

Mode	Prefix	Usage
Query	"Represent this sentence for searching relevant passages: "	User questions via _embed_query()
Document	None (raw text)	Ingested records via to_embedding_text()

This asymmetric approach improves retrieval relevance by 5-15% compared to symmetric encoding.

6.5 Comparative Analysis Mode

Comparative queries are auto-detected and produce structured side-by-side analysis with markdown tables, advantages/limitations, and clinical context.

Detection and Parsing

User Query: "Compare 4-1BB vs CD28 costimulatory domains"
    │
    ├── 1. _is_comparative() detects COMPARE/VS/VERSUS/COMPARING       [< 1 ms]
    │
    ├── 2. _parse_comparison_entities() extracts raw entities            [< 1 ms]
    │      Pattern 1: "X vs/versus Y" (greedy group 2)
    │      Pattern 2: "Compare X and/with Y"
    │      Post-processing: strip trailing context words
    │        (costimulatory domains, resistance mechanisms, etc.)
    │
    ├── 3. resolve_comparison_entity() for each raw entity               [< 1 ms]
    │      Priority: CART_TARGETS → ENTITY_ALIASES → CART_TOXICITIES → CART_MANUFACTURING
    │      "4-1BB" → {"type": "costimulatory", "canonical": "4-1BB (CD137)"}
    │      "CD28"  → {"type": "costimulatory", "canonical": "CD28"}
    │
    ├── 4. Dual retrieve() — one per entity                              [~365 ms]
    │      Entity A: retrieve(question, target_antigen=entity_a.target)
    │      Entity B: retrieve(question, target_antigen=entity_b.target)
    │      ~24 results per entity across 11 collections
    │
    ├── 5. get_comparison_context() — side-by-side knowledge graph       [< 1 ms]
    │      Calls get_target_context() / get_toxicity_context() /
    │      get_manufacturing_context() for each entity
    │
    ├── 6. _build_comparative_prompt() — structured prompt               [< 1 ms]
    │      Evidence grouped by entity (A section, B section)
    │      Knowledge context appended
    │      Instructions: comparison table + advantages + limitations
    │
    └── 7. Stream Claude Sonnet 4.6 (max_tokens=3000)                   [~28-30 sec]
           Structured output: table, pros/cons, clinical context

Total: ~30 sec (365ms retrieval + ~30 sec LLM generation)

Supported Entity Types

Entity Type	Examples	Resolution
Target Antigens	CD19, BCMA, CD22, CD20	Direct match in CART_TARGETS (25 entries)
FDA Products	Kymriah, Yescarta, Carvykti, Abecma	ENTITY_ALIASES → canonical name + target
Costimulatory Domains	4-1BB, CD28	ENTITY_ALIASES → type: costimulatory
Toxicity Profiles	CRS, ICANS, B-cell aplasia	CART_TOXICITIES → type: toxicity
Manufacturing Processes	Lentiviral, transduction	CART_MANUFACTURING → type: manufacturing

Example Comparative Queries

"Compare CD19 vs BCMA"                              → Target vs target (with target_antigen filtering)
"Compare 4-1BB vs CD28 costimulatory domains"        → Costimulatory domain comparison
"Kymriah versus Carvykti"                            → Product vs product (resolves to CD19 vs BCMA)
"Compare CRS and ICANS toxicity"                     → Toxicity profile comparison

Fallback Behavior

If entity parsing fails (unrecognized entities, ambiguous input), the query gracefully falls back to the standard single-query retrieve() path with no user-visible error.

UI Evidence Panel

Comparative evidence is displayed in an entity-grouped collapsible panel:

• Entity A header (blue) — evidence cards for entity A results
• "— VS —" divider (green)
• Entity B header (purple) — evidence cards for entity B results

Each evidence card shows collection badge, ID, cosine score, clickable source link, and text snippet.

6.6 Evidence Panel and Clickable Citations

All query responses include a collapsible evidence panel with collection-badged cards:

Badge Color	Collection	Link Format
Blue	Literature	PubMed
Purple	Trial	ClinicalTrials.gov
Green	Construct	—
Yellow	Assay	—
Orange	Manufacturing	—

Each card displays: collection badge, record ID, cosine similarity score, clickable source link (where available), and a 200-character text snippet.

---

7. Data Sources and Ingest Pipelines

7.1 Actual Ingest Results

Source	API	Records	Time	Collection
PubMed	NCBI E-utilities (esearch + efetch)	5,047	~15 min	cart_literature
ClinicalTrials.gov	REST API v2	973	~3 min	cart_trials
FDA Product Labels	Manual curation (seed script)	6	~5 sec	cart_constructs
Published Assays	Curated JSON (seed script)	45	~30 sec	cart_assays
Published Manufacturing	Curated JSON (seed script)	30	~30 sec	cart_manufacturing
Safety	Curated pharmacovigilance data	40	~30 sec	cart_safety
Biomarkers	Curated CRS/exhaustion markers	43	~30 sec	cart_biomarkers
Regulatory	FDA approval timelines	25	~30 sec	cart_regulatory
Sequences	scFv/molecular binding data	27	~30 sec	cart_sequences
Real-World Evidence	Registry outcomes	30	~30 sec	cart_realworld
Total		6,266	~22 min

7.2 Ingest Pipeline Architecture

All 5 ingest pipelines inherit from BaseIngestPipeline with a standard flow:

fetch(**kwargs)           # Retrieve raw data (API call, file read)
    │
    ▼
parse(raw_data)           # Validate into Pydantic models
    │
    ▼
embed_and_store(records)  # Embed text → insert into Milvus
    │
    ├── record.to_embedding_text()   # Generate embedding input
    ├── embedder.encode(texts)       # BGE-small batch encoding
    ├── Enum → str conversion        # ProcessStep, AssayType, etc.
    ├── UTF-8 byte truncation        # Milvus VARCHAR byte limits
    └── manager.insert_batch()       # Batch insert into collection

7.3 Assay Seed Data Coverage

Category	Records	Key Data
Cytotoxicity	12	All 6 FDA products, CD22, dual-target, GPC3, HER2, Mesothelin, sBCMA decoy
In vivo / Clinical	9	ELIANA, ZUMA-1, ZUMA-2, TRANSFORM, KarMMa, CARTITUDE-1, mouse models
Persistence	5	4-1BB vs CD28 persistence kinetics for all products
Flow cytometry	5	CD19 loss, trogocytosis, lineage switch, BCMA loss, product characterization
Exhaustion	3	CD28 vs 4-1BB exhaustion, ZUMA-1/KarMMa correlative data
Cytokine	3	IFN-gamma profiles: tisagenlecleucel, axicabtagene, lisocabtagene
Proliferation	3	IL-7/IL-15 vs IL-2, tisagenlecleucel 42x, axicabtagene 25x (6-day rapid)

7.4 Manufacturing Seed Data Coverage

Category	Records	Key Data
Transduction	6	Lentiviral titer/VCN/efficiency, retroviral, transposon, mRNA
Expansion	7	IL-2, IL-7/IL-15, rapid 6-day (Kite), defined CD4:CD8 (BMS), POC, T-cell selection, failure modes
Harvest / Formulation	4	Leukapheresis yield, viability, dosing, vein-to-vein time
Cryopreservation	2	Controlled-rate freezing, shipping logistics (LN2 dry shippers)
Release testing	6	Sterility (USP <71>), CAR expression, RCL/RCR, identity/COI, potency, residual beads
Emerging / Cost	5	Allogeneic off-the-shelf, cost breakdown ($150-300K), GMP facility, lymphodepletion, capacity

---

8. Performance Benchmarks

Measured on NVIDIA DGX Spark (GB10 GPU, 128GB unified LPDDR5x memory, 20 ARM cores).

8.1 Ingest Performance

Operation	Time	Records	Rate
PubMed fetch + parse + embed + store	~15 min	5,047	~5.6 rec/sec
ClinicalTrials.gov fetch + embed + store	~3 min	973	~5.4 rec/sec
Assay seed embed + store	~30 sec	45	~1.5 rec/sec
Manufacturing seed embed + store	~30 sec	30	~1.0 rec/sec
FDA construct seed (6 products)	~5 sec	6	~1.2 rec/sec

Note: Ingest rate is dominated by BGE-small embedding time (~180ms per text on CPU). GPU acceleration would increase throughput 10-50x.

8.2 Search Performance

Operation	Latency	Notes
Single collection search (top-5)	3-5 ms	Milvus IVF_FLAT with cached index
11-collection parallel search (top-5 each)	12-16 ms	Sequential per-collection, 55 total results
Query expansion + filtered search	8-12 ms	Up to 5 expanded terms × applicable collections
Knowledge graph augmentation	< 1 ms	In-memory dictionary lookup
Full retrieve() pipeline	20-30 ms	Embed + search + expand + merge + knowledge
Comparative entity resolution	< 1 ms	CART_TARGETS → ENTITY_ALIASES → TOXICITIES → MFG
Comparative dual retrieval (2 × 11 collections)	~365 ms	Two retrieve() calls, ~46 total results (24 + 22)

8.3 RAG Query Performance

Operation	Latency	Notes
Full query() (retrieve + Claude generate)	~24 sec	Dominated by LLM generation
Comparative query (dual retrieve + Claude)	~30 sec	365ms retrieval + structured comparison prompt
Streaming query_stream() (time to first token)	~3 sec	Evidence returned immediately
Response length (standard)	800-2000 chars	Grounded answer with citations
Response length (comparative)	1500-3000 chars	Structured tables + pros/cons + clinical context
Token count (standard)	400-800 tokens	Claude Sonnet 4.6 output
Token count (comparative)	800-1500 tokens	Claude Sonnet 4.6 output (max_tokens=3000)

8.4 Relevance Quality

Query	Top Hit Score	Collection	Expected?
"CD19 CAR-T cytotoxicity"	0.791	cart_assays	Yes (CD22/CD19 assay data)
"BCMA resistance mechanisms"	0.809	cart_assays	Yes (sBCMA decoy mechanism)
"lentiviral VCN transduction"	0.870	cart_manufacturing	Yes (VCN quality attribute)
"CAR-T shipping cryopreservation"	0.779	cart_manufacturing	Yes (cryo shipping logistics)
"CD19 CAR-T failure B-ALL"	0.82-0.90	cart_literature	Yes (PubMed abstracts)

---

9. Infrastructure

9.1 Technology Stack

Component	Technology	Version/Detail
Vector database	Milvus	2.4, localhost:19530
Embedding model	BGE-small-en-v1.5	384-dim, BAAI, ~33M params
LLM	Claude Sonnet 4.6	Anthropic API, claude-sonnet-4-20250514
UI framework	Streamlit	Port 8521, NVIDIA black/green theme
Data models	Pydantic	BaseModel + Field validation
Configuration	Pydantic BaseSettings	Environment variable support
Hardware target	NVIDIA DGX Spark	GB10 GPU, 128GB unified, $3,999

9.2 Service Ports

Port	Service
8521	CAR-T Intelligence Agent Streamlit UI
19530	Milvus vector database (shared with main pipeline)

9.3 Dependencies on HCLS AI Factory

Dependency	Usage
Milvus 2.4 instance	Shared vector database — CAR-T adds 10 owned collections alongside existing genomic_evidence (read-only)
ANTHROPIC_API_KEY	Loaded from rag-chat-pipeline/.env if not set in environment
BGE-small-en-v1.5	Same embedding model as main RAG pipeline

---

10. Demo Scenarios

10.1 Validated Demo Queries

These queries have been tested end-to-end through the RAG pipeline:

1. "Why do CD19 CAR-T therapies fail in relapsed B-ALL patients?" - Searches: literature (resistance papers), trials (terminated CD19 trials), assays (CD19 loss, trogocytosis, lineage switch data), constructs (CD19 product designs) - Knowledge graph: CD19 → known_resistance, toxicity_profile, approved_products - Expected answer: Covers antigen loss (28% of relapses), lineage switch (10%, KMT2A-associated), trogocytosis, T-cell exhaustion

2. "Compare 4-1BB vs CD28 costimulatory domains for DLBCL" - Searches: literature (head-to-head reviews), trials (ZUMA-1 vs TRANSCEND/TRANSFORM), assays (exhaustion markers, persistence data), constructs (Yescarta vs Breyanzi) - Expected answer: CD28 = faster kinetics, higher peak, more exhaustion; 4-1BB = sustained persistence, lower toxicity

3. "What manufacturing parameters predict clinical response?" - Searches: literature (correlative studies), manufacturing (VCN, expansion, phenotype), assays (product characterization), trials (responder analyses) - Expected answer: T-cell fitness, Tcm frequency (>40% threshold), CD4:CD8 ratio, VCN, manufacturing time

4. "BCMA CAR-T resistance mechanisms in multiple myeloma" - Searches: literature (resistance reviews), assays (biallelic BCMA loss, sBCMA decoy), constructs (ide-cel vs cilta-cel), trials (KarMMa/CARTITUDE relapse data) - Expected answer: BCMA downregulation, biallelic loss (29% of relapses), soluble BCMA decoy, antigen density heterogeneity

5. "How does T-cell exhaustion affect CAR-T persistence?" - Searches: literature (exhaustion biology), assays (PD-1/LAG-3/TIM-3 data, CD28 vs 4-1BB exhaustion), manufacturing (IL-7/IL-15 vs IL-2) - Expected answer: Exhaustion markers predict outcomes, CD28 drives faster exhaustion, 4-1BB preserves Tcm, IL-7/IL-15 expansion reduces exhaustion

---

11. File Structure (Actual)

cart_intelligence_agent/
├── Docs/
│   └── CART_Intelligence_Agent_Design.md    # This document
├── src/
│   ├── __init__.py
│   ├── models.py                            # Pydantic data models + ComparativeResult (299 lines)
│   ├── collections.py                       # Milvus collection schemas + manager (842 lines)
│   ├── knowledge.py                         # Knowledge graph + entity aliases + comparison (1,030 lines)
│   ├── query_expansion.py                   # 12 expansion maps, 169→1496 terms (955 lines)
│   ├── rag_engine.py                        # Multi-collection RAG + comparative analysis (558 lines)
│   ├── agent.py                             # CAR-T Intelligence Agent (262 lines)
│   ├── ingest/
│   │   ├── __init__.py
│   │   ├── base.py                          # Base ingest pipeline (184 lines)
│   │   ├── literature_parser.py             # PubMed E-utilities ingest (350 lines)
│   │   ├── clinical_trials_parser.py        # ClinicalTrials.gov API v2 ingest (403 lines)
│   │   ├── construct_parser.py              # CAR construct parser + FDA seed (292 lines)
│   │   ├── assay_parser.py                  # Assay data parser (163 lines)
│   │   └── manufacturing_parser.py          # Manufacturing/CMC parser (111 lines)
│   └── utils/
│       ├── __init__.py
│       └── pubmed_client.py                 # NCBI E-utilities HTTP client (390 lines)
├── app/
│   └── cart_ui.py                           # Streamlit chat + comparative UI (484 lines)
├── config/
│   └── settings.py                          # Pydantic BaseSettings (69 lines)
├── data/
│   ├── reference/
│   │   ├── assay_seed_data.json             # 45 curated assay records
│   │   └── manufacturing_seed_data.json     # 30 curated manufacturing records
│   └── cache/                               # Embedding cache
├── scripts/
│   ├── setup_collections.py                 # Create collections + seed FDA constructs (86 lines)
│   ├── ingest_pubmed.py                     # CLI: PubMed ingest (185 lines)
│   ├── ingest_clinical_trials.py            # CLI: ClinicalTrials.gov ingest (182 lines)
│   ├── seed_assays.py                       # CLI: Seed assay data (95 lines)
│   ├── seed_manufacturing.py                # CLI: Seed manufacturing data (95 lines)
│   ├── validate_e2e.py                      # End-to-end 5-test validation (191 lines)
│   ├── test_rag_pipeline.py                 # Full RAG + LLM integration test (223 lines)
│   └── seed_knowledge.py                    # Knowledge graph export (110 lines)
├── requirements.txt
├── LICENSE                                  # Apache 2.0
└── README.md

55 Python files | ~16,748 lines of code | Apache 2.0

---

12. Implementation Status

Phase	Status	Details
Phase 1: Scaffold	Complete	Data models, collection schemas, knowledge graph, query expansion, RAG engine, agent, Streamlit UI, ingest pipeline stubs
Phase 2: Data Ingest	Complete	PubMed (5,047), ClinicalTrials.gov (973), FDA constructs (6), assay seed (45), manufacturing seed (30), safety (40), biomarkers (43), regulatory (25), sequences (27), real-world (30)
Phase 3: RAG Integration	Complete	Multi-collection search, knowledge augmentation, Claude Sonnet 4.6 streaming, 5 demo queries validated
Phase 4: UI + Demo	Complete	Streamlit UI on port 8521, NVIDIA theme, sidebar filters, demo query buttons, streaming responses
Phase 5: UI + Analysis	Complete	Clickable PubMed/ClinicalTrials.gov citation links, collapsible evidence panel with collection badges, Comparative Analysis Mode with auto-detection, entity resolution (39+ aliases), dual retrieval (~365ms), entity-grouped evidence panel, and structured markdown comparison tables
Phase 6: Report Export	Complete	PDF, Markdown, and JSON export via src/export.py (904 lines). NVIDIA-themed PDF with reportlab Platypus: collection-specific evidence tables, clickable PubMed/ClinicalTrials.gov citations, markdown-to-PDF conversion, comparative report support. Download buttons in Streamlit UI after every query response.

Remaining Work

Item	Priority	Effort
Agent reasoning loop testing (agent.py plan→search→synthesize)	Medium	1-2 hours
Genomic evidence bridge (connect to existing genomic_evidence collection)	Low	2-3 hours
Nextflow orchestrator integration	Low	1-2 hours
Additional construct data (published designs beyond FDA-approved)	Low	2-3 hours
Landing page integration (health check endpoint)	Low	1 hour

---

13. Relationship to HCLS AI Factory

This agent demonstrates the generalizability of the HCLS AI Factory architecture. The same infrastructure that supports the VCP/Frontotemporal Dementia pipeline extends to CAR-T cell therapy intelligence with:

• Same Milvus instance — 10 new owned collections alongside existing genomic_evidence (3.56M vectors, read-only)
• Same embedding model — BGE-small-en-v1.5 (384-dim)
• Same LLM — Claude via Anthropic API
• Same hardware — NVIDIA DGX Spark ($3,999)
• Same patterns — Pydantic models, BaseIngestPipeline, knowledge graph, query expansion

The key architectural insight: the platform is not disease-specific. By changing the knowledge graph, query expansion maps, and collection schemas, the same RAG architecture serves any therapeutic area. The CAR-T agent proves this with a completely different domain (cell therapy manufacturing vs. small molecule drug discovery) running on the same infrastructure.

---

14. Credits

• Adam Jones — HCLS AI Factory platform, 14+ years genomic research
• Apache 2.0 License