Indra — Epistemic Research Landscape Agent

Indra is an evidence-backed research navigator for scientific literature. It searches academic sources, grows branching research sessions, validates atomic claims against inspectable evidence, maps literature, surfaces uncertainty, and exports reusable research artifacts.

Indra is not primarily a generic chatbot or writing assistant. Its product surface is a research mission-control workspace for understanding a field, inspecting evidence, finding uncertainty, and leaving with useful artifacts.

Implemented product phases

• Academic search through Semantic Scholar and arXiv.
• Composite multi-provider search with parallel, fallback, and single-source strategies.
• PDF text extraction and OpenRouter-compatible summarization.
• Recursive research orchestration with branches, loops, reflection, and hypothesis generation.
• FastAPI product API with in-memory and Postgres repositories.
• Durable background-job contracts and worker leasing primitives.
• Next.js project and session dashboard.
• Atomic claim extraction, evidence retrieval, claim validation, and claim inspection.
• Citation/reference research maps, timelines, clusters, paper roles, and related-paper recommendations.
• Contradiction, weak-evidence, gap, open-problem, recommendation, and speculative-hypothesis analysis.
• Phase 8 exports: BibTeX, RIS, Markdown report, LaTeX outline, annotated bibliography, claim-ledger CSV/JSON, and research-map JSON.

Architecture

apps/web/                        Next.js product dashboard
src/api/                         FastAPI product API
src/claims/                      claim extraction, evidence retrieval, validation
src/maps/                        research-map construction
src/analysis/                    gap, contradiction, and advisor analysis
src/exports/                     deterministic research artifact generation
src/jobs/                        durable worker boundary
src/                             research core and provider integrations
migrations/                      Postgres schema migrations

Runtime boundary:

frontend -> product API -> repository/events
                       -> durable jobs/workers -> research core -> providers

The frontend does not import the Python research core. Long-running research work must remain behind the durable job boundary.

Installation

Python 3.13 or newer is required.

uv sync

Alternative:

python -m venv .venv
source .venv/bin/activate
pip install -e .

Environment variables

Create a local .env file. Do not commit secrets.

OPENROUTER_API_KEY=...
OPENROUTER_BASE_URL=https://openrouter.ai/api/v1
OPENROUTER_MODEL=anthropic/claude-3-5-sonnet

INDRA_REPOSITORY_BACKEND=memory
INDRA_DATABASE_URL=postgresql://user:password@localhost:5432/indra
INDRA_CORS_ORIGINS=http://localhost:3000,http://127.0.0.1:3000

SEMANTIC_SCHOLAR_API_KEY=...
HALUGATE_URL=http://localhost:8000

Run the API

uvicorn src.api.app:app --reload --port 8000

Major endpoints:

• POST /projects, GET /projects, GET /projects/{project_id}
• POST /sessions, GET /sessions, GET /sessions/{session_id}
• GET /sessions/{session_id}/state
• GET /sessions/{session_id}/map
• GET /sessions/{session_id}/analysis
• GET /sessions/{session_id}/exports
• GET /sessions/{session_id}/exports/{format_name}
• POST /sessions/{session_id}/start|pause|resume|cancel
• GET /sessions/{session_id}/branches|papers|claims|jobs|events
• GET /sessions/{session_id}/events/stream
• POST /branches/{branch_id}/continue|split|prune
• GET /papers/{paper_id}
• POST /sessions/{session_id}/claims/extract
• POST /claims/{claim_id}/validate
• POST /claims/{claim_id}/validate/auto
• GET /claims/{claim_id}/inspection
• POST /jobs/lease, POST /jobs/{job_id}/complete|fail

The default repository backend is process-local memory. Set INDRA_REPOSITORY_BACKEND=postgres and INDRA_DATABASE_URL for durable persistence.

Run the dashboard

cd apps/web
cp .env.example .env.local
npm install
npm run dev

Open http://localhost:3000/projects.

The dashboard includes:

• project and session workspaces;
• session lifecycle controls and live events;
• branch, paper, job, claim, and evidence inspectors;
• research maps and timelines;
• research-advisor recommendations, contradiction and gap review, hypotheses, and weak-evidence triage;
• export center at /sessions/{session_id}/exports.

Phase 8 export formats

Format	Endpoint suffix
BibTeX	bibtex
RIS	ris
Markdown research report	report-markdown
LaTeX literature-review outline	literature-review-latex
Annotated bibliography	annotated-bibliography
Claim ledger CSV	claim-ledger-csv
Claim ledger JSON	claim-ledger-json
Research map JSON	research-map-json

Claim-bearing exports preserve status, confidence, evidence relationships, and synthesis eligibility. Unsupported, contradicted, speculative, and unreviewed statements remain explicitly labelled.

Verification

Frontend:

cd apps/web
npm run typecheck
npm test
npm run build

Backend:

python -m pytest -q \
  test_api_cors.py \
  test_api.py \
  test_claim_evidence_retrieval.py \
  test_research_map.py \
  test_research_advice.py \
  test_exports.py

Implementation notes:

• docs/phase4/WEB_DASHBOARD_MVP.md
• docs/phase5/CLAIM_VALIDATION_MVP.md
• docs/phase6/RESEARCH_MAPS_MVP.md
• docs/phase7/RESEARCH_ADVISOR_MVP.md
• docs/phase8/EXPORTS_MVP.md

Production-hardening work still required

• Connect durable research jobs to full MasterAgent execution.
• Replace process-local SSE with resumable cross-process events.
• Add authentication and project authorization.
• Expose full-text paper chunks to evidence retrieval.
• Add calibrated domain-specific inference where appropriate.
• Deploy Postgres, migrations, API, workers, and dashboard as one system.
• Add large-session caching and asynchronous export jobs if session scale requires them.

Core rule

Factual claims must be decomposed, validated, and linked to source evidence. Unsupported output remains excluded or explicitly uncertain. Hypotheses remain speculative until independently evidenced. Exports must preserve these distinctions.