delhpi

<pre align="center"> ██████╗ ███████╗██╗ ██████╗ ██╗ ██╗██╗ ██╔══██╗██╔════╝██║ ██╔══██╗██║ ██║██║ ██║ ██║█████╗ ██║ ██████╔╝███████║██║ ██║ ██║██╔══╝ ██║ ██╔═══╝ ██╔══██║██║ ██████╔╝███████╗███████╗██║ ██║ ██║██║ ╚═════╝ ╚══════╝╚══════╝╚═╝ ╚═╝ ╚═╝╚═╝ </pre> <h3 align="center">open-source MCP server for AI agents</h3> <p align="center">Index code repositories, research papers, and HuggingFace datasets. Runs fully local.</p> <p align="center"> <a href="https://www.apache.org/licenses/LICENSE-2.0"><img src="https://img.shields.io/badge/License-Apache_2.0-orange.svg" alt="License"></a> <a href="https://www.python.org/"><img src="https://img.shields.io/badge/python-3.10+-orange.svg" alt="Python"></a> <a href="https://modelcontextprotocol.io/"><img src="https://img.shields.io/badge/MCP-compatible-orange.svg" alt="MCP"></a> </p> --- ## What is Delphi? Delphi is a self-hosted [MCP (Model Context Protocol)](https://modelcontextprotocol.io/) server that gives AI agents deep context through semantic search across three domains: | Domain | What it does | |--------|-------------| | **Code Repositories** | Index GitHub repos, search code semantically, find symbols (functions, classes), analyze architecture | | **Research Papers** | Index arXiv papers or PDFs, extract citations and equations, generate reports | | **HuggingFace Datasets** | Index dataset cards, search metadata | Everything runs on your machine — PostgreSQL for storage. Pick **local sentence-transformers** (no API keys) or wire up **Gemini** / **OpenAI** for hosted embeddings. --- ## Latest results On [Agent Retrieval Bench v2](https://github.com/eyuansu62/agent-retrieval-bench) (commit `d04953371d96`), Delphi leads every published baseline on MRR and Recall@5 over the same 75 development cases, with the same `all_files` candidate filter, scored by the benchmark's own code. | System | MRR | Recall@5 | Recall@20 | Latency | | --- | ---: | ---: | ---: | ---: | | **Delphi** | 0.220 | 0.369 | **0.551** | **5.7 s** | | Nia (hosted) | 0.228 | **0.391** | 0.449 | 36.9 s | | grep | 0.180 | 0.302 | 0.578 | — | | RepoMap | 0.169 | 0.240 | 0.551 | — | | lexical | 0.127 | 0.198 | 0.451 | — | | BM25 | 0.116 | 0.136 | 0.429 | — | Every row above is the 75-case development split, because that is the split ARB's published baselines were run on; mixing sample sizes inside one table would not be a comparison. The Delphi-versus-Nia question is answered separately, on the 135 cases where both engines have the corpus indexed — a 0.02 difference cannot be resolved at n=75 when run-to-run variance is itself about 0.02: | Metric | Delphi | Nia | | --- | ---: | ---: | | MRR | 0.229 | **0.261** | | Recall@5 | 0.350 | 0.360 | | Recall@20 | **0.528** | 0.424 | Zero failures on either side. Counting per-case outcomes rather than averages: | Metric | Delphi wins | Nia wins | Ties | | --- | ---: | ---: | ---: | | MRR | 40 | 47 | 48 | | Recall@5 | 21 | 22 | 92 | | Recall@20 | **31** | 12 | 92 | **Nia ranks the top of the list better.** It leads MRR by 0.031 and wins more cases head to head. Recall@5 is a genuine tie, 21 cases to 22. What Delphi wins is coverage — Recall@20 by 0.104, and 31 cases to 12 — and latency, by about 6.5x, on your own hardware. We ran the larger comparison specifically to test a more flattering number from the smaller one, and it did not survive. grep still leads Recall@20 outright at 0.578. The two shapes follow from different strategies rather than different amounts of skill: Nia returns ~7.8 files per query, Delphi returns 20. A short confident list wins precision at the top; a longer one wins coverage. ### What changed, and what each change was worth The previous evaluation measured a corpus indexed with `text-embedding-3-small` but queried with `gemini-embedding-001`. Both emit 768-dimensional vectors, so pgvector computed cosines between unrelated spaces without raising anything. Embedding a chunk's own exact content and comparing it against that chunk's stored vector scored **cosine 0.0101** — orthogonal. Aligning the query-time model to the index moved the same comparison to **0.9430**. | Configuration | MRR | Recall@5 | Recall@20 | Latency | | --- | ---: | ---: | ---: | ---: | | As previously benchmarked | 0.055 | 0.056 | 0.161 | 1.0 s | | Embedding space aligned | 0.173 | 0.259 | 0.549 | 0.9 s | | + rank fusion, path affinity | 0.176 | 0.256 | 0.552 | 1.1 s | | + cross-encoder rerank | **0.193** | **0.294** | **0.560** | 2.9 s | Delphi now compares `repositories.embedding_model` against the model answering queries on every search and on `/backend-health`, so a silent vector-space mismatch is reported instead of absorbed. ### Held-out results, by workflow All 220 positive cases of the final split, zero failed queries: | Workflow | Cases | MRR | Recall@5 | Recall@20 | | --- | ---: | ---: | ---: | ---: | | trace2code | 38 | 0.693 | 0.842 | 0.908 | | edit2ripple | 44 | 0.274 | 0.434 | 0.587 | | code2test | 83 | 0.220 | 0.394 | 0.586 | | comment2context | 55 | 0.207 | 0.245 | 0.345 | | **overall** | **220** | **0.309** | **0.442** | **0.582** | Queries about a code index are usually written in English while the index is written in code, so Delphi can embed a hypothetical code snippet alongside the question (`SYNSC_QUERY_EXPANSION=true`). On the held-out split that moves every metric: | Configuration | MRR | Recall@5 | Recall@20 | Latency | | --- | ---: | ---: | ---: | ---: | | Retrieval only | 0.228 | 0.349 | 0.552 | 1.96 s | | + hypothetical document | 0.241 | 0.355 | **0.579** | 4.11 s | | + listwise rerank | **0.309** | **0.442** | **0.582** | 5.57 s | The spread matters more than the average. A failure trace hands the retriever real symbols and stack frames, and Delphi finds the root-cause file in the top five 74% of the time. A review comment hands it English, and the same engine manages 15%. That gap is the difference between a query that contains evidence and one that does not. ### Downstream: a null result Retrieval metrics measure whether the right file was found. DS-1000 measures whether the agent's generated code passes the official test, with the model held fixed — an outcome a retrieval benchmark cannot fudge. On 40 development tasks Delphi reached 0.900 against 0.875 for both hosted engines, which reads like a downstream lead. It is not. On 100 held-out tasks, against a no-retrieval control: | Condition | pass@1 | | --- | ---: | | No retrieval | **0.870** | | Delphi | 0.860 | Paired per case: 2 tasks Delphi passed and the control did not, 3 the other way, and **95 of 100 decided identically**. Documentation context does not change what this model produces on this benchmark in either direction, and the 40-case spread was one or two tasks of noise. We report it because a benchmark that cannot separate the conditions is worth saying out loud — particularly when its smaller slice flatters us. Anyone claiming a large DS-1000 retrieval win on 40 tasks is measuring sampling error. ### Scope and limits - The held-out split is reported at full scope: all 220 positive cases, every corpus provisioned, zero failed queries. Delphi scores 0.309 MRR / 0.442 Recall@5 / 0.582 Recall@20 with query expansion and listwise reranking enabled — well ahead of the development split the pipeline was tuned on. - The hosted head-to-head is run on the development split, where both engines have every corpus indexed. Nia has 60 of the 220 final-split commits indexed on its side, so a full held-out head-to-head is not available. - Delphi is not state of the art at the top of the list. On 135 shared cases the hosted comparator leads MRR 0.261 to 0.229 and wins more cases head to head. Recall@5 is tied. Delphi's wins are coverage and latency. - Every reported configuration was confirmed on the 220-case held-out split before shipping. Four candidate improvements looked good on the development split and were rejected when held-out disagreed. - Earlier head-to-head figures are withdrawn rather than restated: the Delphi half of that run is now known to have been measuring a mismatched embedding space. Full method, ablations, and expandable per-query retrieval traces: [The context engine is the product](https://trydelphi.ai/blog/context-engine-is-the-product). ### Timeline | When | Milestone | | --- | --- | | Open source | Released Delphi as a local-first MCP context engine under Apache 2.0. | | Product foundation | Added versioned multi-source indexing, hybrid retrieval, code intelligence, and agent-ready context packs. | | 2026-07-30 | Found and fixed a silent embedding-space mismatch that had made the vector branch return random results; added a permanent check for it. | | 2026-07-30 | Rebuilt ranking on reciprocal-rank fusion, added a path-affinity branch, and set reranker defaults from measurement. | --- ## Quick Start ```bash npx @synsci/delphi ``` That's it. The installer asks two questions: 1. **Add Delphi to your coding agent** (Claude Code, Cursor, Windsurf, Claude Desktop) — or **run your own index** with a Gemini / OpenAI / local-model key and a dashboard. 2. Which embeddings provider to use. Then it pulls the source, spins up the Docker stack, mints an API key, and (if you picked the agent path) writes the MCP config for the tools you have installed. When it finishes, restart your AI tool — Delphi shows up as an MCP server. After install, just type `delphi` in any terminal to open the dashboard. ```bash delphi # open the dashboard (boots the stack if it's down) delphi status # check health + container state delphi logs -f # tail logs delphi stop # tear it down delphi uninstall # remove containers + data volume ``` > Requires Docker Desktop (or `docker compose` v2) and `git`. > **Dashboard:** [localhost:3000](http://localhost:3000) &nbsp;&middot;&nbsp; **API:** [localhost:8742](http://localhost:8742) <details> <summary><strong>Manual install (from source)</strong></summary> For contributors or anyone who wants to run a fork: ```bash git clone https://github.com/synthetic-sciences/delphi.git cd delphi cp env.example .env # set SERVER_SECRET and SYSTEM_PASSWORD ./scripts/launch_app.sh # or: docker compose up --build ``` </details> <details> <summary><strong>Manual MCP config (for any client the installer doesn't cover)</strong></summary> Once you have an API key (from the dashboard at `/api-keys` or via `npx @synsci/delphi init`), add this to your client's MCP config: ```json { "mcpServers": { "synsci-delphi": { "command": "uvx", "args": ["synsci-delphi-proxy"], "env": { "SYNSC_API_KEY": "your-api-key", "SYNSC_API_URL": "http://localhost:8742" } } } } ``` **Config file paths:** Cursor `~/.cursor/mcp.json` · Windsurf `~/.codeium/windsurf/mcp_config.json` · Claude Desktop `~/Library/Application Support/Claude/claude_desktop_config.json` · Claude Code use `claude mcp add --scope user synsci-delphi -- uvx synsci-delphi-proxy`. </details> <details> <summary><strong>HTTP API (any client)</strong></summary> ```bash curl http://localhost:8742/health # Index a repository curl -X POST http://localhost:8742/api/repositories/index \ -H "Authorization: Bearer your-api-key" \ -H "Content-Type: application/json" \ -d '{"url": "https://github.com/owner/repo"}' # Search code curl "http://localhost:8742/api/search/code?query=authentication+middleware" \ -H "Authorization: Bearer your-api-key" ``` </details> ### Reproducible source versions Completed indexing runs publish immutable, content-addressed snapshots of repositories, papers, datasets, and documentation. Each version keeps its own normalized chunks and vectors, while a separate head identifies the current version. This is fully local and does not require a hosted service. ```bash # Authenticated HTTP curl "http://localhost:8742/v2/sources/SOURCE_ID/snapshots?type=repo" \ -H "Authorization: Bearer your-api-key" curl "http://localhost:8742/v2/snapshots/SNAPSHOT_ID?include_items=true" \ -H "Authorization: Bearer your-api-key" # Local database CLI synsc-context snapshots list --type repo --source-id SOURCE_ID --json synsc-context snapshots show SNAPSHOT_ID --include-items --json ``` ### Policy-aware retrieval Retrieval is planned before providers run. The planner searches the local index first, keeps pinned snapshot scopes exact, and only admits optional remote search when both the deployment network ceiling and the request-level egress policy allow it. Plans are deterministic and auditable: admitted steps, denied alternatives, provider choices, privacy decisions, and hard call/result/ provider-hit-payload/deadline budgets are recorded without exposing user identities. Execution binds the plan to the independently authenticated user and revalidates plan integrity, provider health and capabilities, and egress immediately before each call. Providers receive the remaining deadline, a bounded response contract, and cooperative cancellation. Provider failures remain isolated, and results carry fused per-provider provenance. The default remains local-only; remote providers are optional. An optional hosted adapter supplies web search and bounded public-site crawling without changing the local-first default. It is registered lazily, requires no cloud credential at startup, and cannot run unless both deployment policy and the individual request permit public egress. See [`docs/env-advanced.md`](docs/env-advanced.md#optional-provider-network-policy). ### Incremental connectors Connectors ingest changing sources through a durable, local-first sync queue. Configuration and checkpoints are encrypted, changes are bounded and idempotent, deletions and permission revocations become tombstones, and a cursor advances only in the transaction that activates its validated immutable snapshot. The built-in `local-folder` adapter never uses the network; hosted adapters remain optional. Local-folder access is denied until the operator sets `SYNSC_LOCAL_CONNECTOR_ALLOWED_ROOTS`. See [`docs/connectors.md`](docs/connectors.md). ### Reproducible context sessions Context sessions keep task state, pinned snapshot references, accepted and rejected evidence, decisions, unresolved questions, and deterministic token-budget manifests in append-only revisions. Parent/child handoffs preserve the exact revision they continue from. Reads and exports re-check current source authorization, so a later permission revocation does not leak content through a saved context. See [`docs/context-sessions.md`](docs/context-sessions.md). ### One workspace, three interfaces The dashboard's `/workspace` route brings provider health, connector sources, durable research, and reproducible contexts into one local-first control plane. The same safe surface is available to scripts through the Python client and CLI: ```bash export SYNSC_API_URL=http://localhost:8742 export SYNSC_API_KEY=your-api-key synsc-context workspace synsc-context connectors list synsc-context connectors sync SOURCE_ID synsc-context contexts create release-review \ --objective "Verify the release against pinned evidence" \ --snapshot-id SNAPSHOT_ID synsc-context contexts export SESSION_ID --json ``` ```python from synsc.client import SynscClient with SynscClient() as client: workspace = client.workspace() ``` Authentication stays in environment variables; neither the dashboard nor these commands returns connector configuration or provider credentials. --- ## MCP Tools > **MCP defaults to `quality_mode='agent'`.** Indexing includes tests, docs, examples, configs, manifests, and dotfiles. Search runs hybrid retrieval (vector + BM25 + exact symbol + exact path + trigram) with stable file-level diversity. Structured developer requests with safe repository-relative file fields also probe related paths (for example, `core_model_loading.py` can surface `test_core_model_loading.py`). Deployments can opt into a cross-encoder rerank with `SYNSC_ENABLE_RERANKER=true`. Each result carries `candidate_sources` so the agent can see which branches surfaced it. Pass `quality_mode='fast'` for the legacy pure-vector path. ### Code | Tool | Description | |------|-------------| | `index_repository` | Index a GitHub repository. Accepts `quality_mode`, `include_tests`, `include_docs`, `include_examples`, `deep_index`, `force_reindex`. Branch is optional — the default branch is auto-detected. | | `index_local_folder` | Index a **local directory** straight from disk (no GitHub needed) — private/work code and work-in-progress. | | `quick_index` | Resolve a library name via the curated catalog and index it in one step (`quick_index("fastapi")`). | | `catalog_search` | Resolve a library/framework name to an indexable source — works with **zero** indexed sources (cold start). | | `search_code` | Hybrid code search (vector + BM25 + symbol + path + trigram, fused). | | `search_symbols` | Find functions, classes, methods by name. Structural extraction for **16 languages** (Python, JS/TS, Go, Rust, Java, C, C++, C#, Ruby, PHP + Kotlin/Swift/Scala/Lua/Elixir/shell via regex). | | `get_symbol` | Full symbol details *plus the reconstructed source body* (no separate `get_file` call needed). | | `find_callers` | Who calls this symbol? (code-dependency graph) | | `find_callees` | What does this symbol call — internal symbols + external names. | | `impact_analysis` | Blast radius — transitive callers: "what breaks if I change this function?" | | `build_code_graph` | (Re)build the call graph for a repo (built automatically after indexing). | | `check_freshness` | Is this index stale? Compares against remote HEAD (git) or re-hashes files (local). | | `list_stale_sources` | List indexed repos whose index has drifted from its source. | | `get_file` | Retrieve file content from an indexed repo. | | `get_context` | Fetch a chunk plus adjacent chunks, enclosing function/class body, and same-class siblings. | | `build_context_pack` | Agent-ready pack: primary hits + enclosing bodies + adjacent chunks + same-class siblings + imports + linked tests/docs/examples/configs + symbol details + architecture summary, with a token budgeter and a re-query planner. | | `get_directory_structure` | Browse repository file tree. | | `analyze_repository` | Deep code analysis and architecture overview. | | `classify_failure` | Tag a "Delphi failed because" event with a stable failure-mode code. | > **Code intelligence:** after indexing, Delphi builds a symbol-level call graph so agents can reason about structure (`find_callers`, `impact_analysis`), not just retrieve text. **Cold start:** a curated catalog maps popular library names to sources so `catalog_search`/`quick_index` work before you've indexed anything. **Freshness:** `check_freshness` flags stale indexes. See [`docs/cold-start.md`](docs/cold-start.md). ### Papers | Tool | Description | |------|-------------| | `index_paper` | Index from arXiv URL/ID or PDF upload. | | `search_papers` | Hybrid paper search with section-aware (Methods > Related Work) and citation-aware ranking, cross-encoder rerank. | | `extract_quoted_evidence` | Pull literal sentences from a paper that ground a claim. | | `joint_retrieval` | One call → paper + code + Atlas-graph hits, fused. | | `get_citations` | Extract citation graph. | | `get_equations` | Extract equations with context. | | `generate_report` | Generate a markdown summary report. | | `compare_papers` | Side-by-side paper comparison. | ### Datasets | Tool | Description | |------|-------------| | `index_dataset` | Index a HuggingFace dataset card. | | `search_datasets` | Semantic dataset search with cross-encoder rerank. | ### Durable research | Tool | Description | |------|-------------| | `research_start` | Queue a `quick`, `deep`, or `oracle` research session and return immediately. | | `research_list` | List the caller's sessions, including work recovered after a restart. | | `research_status` | Read the latest persisted answer, citations, usage, and status. | | `research_events` | Replay the append-only progress log from a sequence cursor. | | `research_followup` | Persist a follow-up turn and requeue the same conversation. | | `research_cancel` | Cancel pending work or request cooperative cancellation of a running job. | Async research is backed by PostgreSQL rather than API-process memory. The worker uses leased claims, heartbeats, bounded retries, stale-job recovery, and generation fencing so an interrupted or superseded worker cannot publish a late answer. The HTTP equivalents are under `/v2/research`; SSE reconnects can send `Last-Event-ID` and receive only later persisted events. ### Reproducible contexts | Tool | Description | |------|-------------| | `context_session_create` | Create a private-by-default session from pinned snapshot references. | | `context_session_list` | List the caller's current context-session metadata. | | `context_session_get` | Rehydrate an authorized current or historical revision. | | `context_session_revise` | Append an immutable revision behind an optimistic write fence. | | `context_session_handoff` | Create an explicitly linked child from the current parent revision. | These tools are included in the `all`, `code`, `papers`, `docs`, and `atlas` profiles. The compact `minimal` profile omits them. ### Atlas integration (optional) Delphi can ingest and retrieve over an [Atlas](https://github.com/synthetic-sciences/atlas) research graph (nodes, edges, artifacts, executions, tool contracts), with graph-aware context packs and "what was tried / don't-repeat / decision recall" surfaces. **Off by default** — these tools only make sense if you're pushing graph data into Delphi from an Atlas workspace, and they cost MCP-handshake tokens when exposed. Set `SYNSC_MCP_PROFILE=atlas` (or `all`) to turn them on. Full tool list and ingestion contracts in [`docs/atlas-integration.md`](docs/atlas-integration.md). --- ## Architecture ``` ┌──────────────────────────────────────────────┐ │ AI Agent (Claude, Cursor, etc.) │ │ Calls MCP tools to index & search │ └──────────────┬───────────────────────────────┘ │ MCP (stdio) or HTTP ┌──────────────▼───────────────────────────────┐ │ Delphi Server (FastAPI) │ │ ┌──────────┐ ┌───────────┐ ┌─────────────┐ │ │ │ Indexing │ │ Search │ │ Papers │ │ │ │ Service │ │ Service │ │ Service │ │ │ └─────┬────┘ └─────┬─────┘ └──────┬──────┘ │ │ ┌─────▼─────────────▼──────────────▼──────┐ │ │ │ sentence-transformers (local) │ │ │ │ No API keys needed │ │ │ └─────────────────────────────────────────┘ │ └──────────────┬───────────────────────────────┘ │ ┌──────────────▼───────────────────────────────┐ │ PostgreSQL + pgvector │ │ All data stays on your machine │ └──────────────────────────────────────────────┘ ``` --- ## Project Structure ``` backend/ Python backend (FastAPI + MCP) synsc/ Application package api/ HTTP + MCP server entry points services/ Business logic (search, indexing, papers, datasets) database/ SQLAlchemy models, session management embeddings/ sentence-transformers embedding provider extractors/ Symbol extraction (tree-sitter AST) indexing/ Repo/paper/dataset indexing pipelines parsing/ Language parsers workers/ Background indexing worker alembic/ DB migrations tests/ Pytest suite pyproject.toml Python deps & entry points Dockerfile Image for api + worker targets frontend/ Next.js dashboard landing/ Public Next.js site (deployed from this monorepo) packages/cli/ `npx @synsci/delphi` installer (one-command setup) packages/mcp-proxy/ MCP stdio-to-HTTP bridge (published separately) database/supabase/ Local PostgreSQL init SQL scripts/ Developer scripts (launch_app.sh, etc.) docs/ Architecture + engineering docs docker-compose.yml Local dev stack (postgres + api + worker + frontend) ``` See [`docs/architecture.md`](docs/architecture.md) for a walk-through of how the pieces fit together. ## Configuration All configuration is via environment variables. See [`env.example`](env.example) for the full list. | Variable | Default | Description | |----------|---------|-------------| | `DATABASE_URL` | `postgresql://synsc:synsc@localhost:5432/synsc` | PostgreSQL connection | | `SERVER_SECRET` | — | JWT signing secret (required) | | `SYSTEM_PASSWORD` | — | Admin login password | | `EMBEDDING_MODEL` | `all-mpnet-base-v2` | sentence-transformers model | | `EMBEDDING_DEVICE` | auto | `cpu`, `cuda`, or `mps` | | `SYNSC_ENABLE_RERANKER` | `false` | Enable cross-encoder reranking | | `HF_TOKEN` | — | HuggingFace token for dataset indexing | ## Lite deployment For a laptop, CI, or air-gapped box, the **lite stack** (Postgres + API only) skips the ~1.2 GB embedding-model download via `EMBEDDING_PROVIDER=hash` and boots in seconds — hybrid retrieval's lexical + symbol branches still carry most of the quality. See [`docs/deployment-lite.md`](docs/deployment-lite.md). ```bash docker compose -f docker-compose.lite.yml up --build ``` ## Benchmark Retrieval quality and token economy are measured, not asserted. The reproducible, database-free harness in [`bench/`](bench/) scores recall@k, precision@k, nDCG, MRR, and token cost across naive grep, smart grep, BM25, symbol lookup, and Delphi's hybrid fusion: ```bash uv run --project backend python bench/run.py ``` Point it at your own corpus with `--corpus` / `--tasks`. Methodology and sample numbers are in [`bench/README.md`](bench/README.md). ## Development All Python commands run from `backend/`: ```bash cd backend uv run pytest # tests uv run ruff check synsc/ tests/ # lint uv run ruff format synsc/ tests/ # format uv run mypy synsc/ # type check ``` --- ## License [Apache License 2.0](LICENSE)